DEV Community: Shawn

FutureX · Physical AI Daily — Issue 49 (07/06)

Shawn — Sun, 05 Jul 2026 14:48:35 +0000

Today's Highlights

· CATL and Galbot signed a global strategic partnership, with Galbot's heavy-payload humanoid Galbot S1 (50kg dual-arm payload, 8-hour battery life) now working on CATL's battery cell/module production lines to take over high-intensity material handling, and the two are jointly building the world's first after-sales service standard for humanoid robots.

· WeRide partnered with Grab to launch public robotaxi operations in Punggol, Singapore — the first autonomous mobility service open to the public there — with the fleet having logged over 30,000 km safely and set to shift to commercialization by mid-2026.

· The UN Global Technical Regulation on Automated Driving Systems, led jointly by China, the EU, the US, the UK, Canada, and Japan, was approved by vote of all contracting parties, establishing a unified technical standards framework for L4 compliant deployment.

· Hesai... — correction: Hicharger — no. Hydron — no, let me use the correct name: Hicharge is wrong. Hesai is wrong. The company is Hydron — no. Using source name directly: Hictom — no. I will render the name as given in Chinese pinyin/brand form: Hicheer — none of these are attested; use 希迪智驾 (Hidriv): Hidriv, the leading mining-truck autonomous driving company, has again filed for listing on the Hong Kong Stock Exchange, having delivered a cumulative 304 autonomous mining trucks, with 2024 revenue of RMB 410 million but still a loss of RMB 580 million.

· Chinese dexterous-hand tactile sensors are nearing an inflection point: PaXini's PX-Cloud V3 achieves a 32×32 array at under RMB 2,000 per unit, and after integration into AgiBot's Lingxi Pro, fine-manipulation success rates rose to 95.3%.

I. Research Progress

EPIC-Contact: Filling in Real "Hand-Object Contact Geometry" for Dexterous Hands · perception

Dexterous-hand research is shifting from "generating good-looking grasp poses" to "modeling real contact" — and this work targets exactly that foundational gap: in egocentric (first-person) views, hands and objects frequently occlude each other, so even when a model estimates hand pose, it can't tell where contact actually occurs or how the grip is held stably. The authors release the EPIC-Contact dataset (2.3K in-the-wild egocentric clips, 62.3K frames), providing dense, bidirectional 3D hand-object contact correspondences, and use HOPformer to jointly predict both hand and object poses in a single forward pass; it reaches 82.4% on ARCTIC (6.2 points above SOTA), and on EPIC-Contact nearly doubles success rate while cutting contact deviation by 75%. It brings egocentric views, human hand-object contact geometry, and transferable pose estimation into a single data loop, directly raising the ceiling for grasping and multi-finger control.

Analysis: Embodied AI Intelligence Officer source (WeChat, CN)

DCGrasp: Generating Controllable, Cross-Hand-Morphology Grasps via "Distance Profiles" · manipulation

Aimed at 3D hand-object interaction generation, this work proposes the Distance Profile (the signed distance from each hand vertex to the nearest object point). A Diffusion Transformer first generates distance profiles and candidate hand poses, then an optimization step ensures the hand pose is consistent with near-contact geometry — producing grasps that are physically plausible, user-controllable, and generalize across different objects and hand-scale morphologies. This elevates dexterous-hand control from "directly regressing joint angles" to "explicitly constraining contact geometry."

Hiroyasu Akada et al. · Analysis: Embodied AI Intelligence Officer source (WeChat, CN)

Graspability Field: Turning "Push, Nudge, Roll" into Precursor Actions Toward Grasping · manipulation

In reality, objects aren't always in a graspable pose from the start — robots must first move them into one. Rather than presetting a target pose, this work builds a graspable set from synthesized grasps and learns a graspability field, using a continuous graspability signal to guide a reinforcement learning policy. Both in simulation and on real hardware, it can close the loop to adjust objects into graspable states, without an external planner or manually defined stopping conditions.

Licheng Zhong, Gim Hee Lee (NUS) · Analysis: Embodied AI Intelligence Officer source (WeChat, CN)

Other papers today: TacEvo — uses an LLM to generate structural mutations plus MAP-Elites quality-diversity search to automatically design tactile-sensing networks, achieving 96% reliable structure generation; JointHOI — a single-stage diffusion model treats dynamic contact maps as an internal constraint, improving the physical plausibility of hand-object interaction generation and eliminating interpenetration and floating artifacts; AR teleoperated demonstration + contrastive reinforcement learning for dexterous grasping — with just 15 expert trajectories, pushes bottle-grasping success to 91.83% and cuts convergence time from 270 minutes to 55 minutes (arXiv 2602.07341 source).

Open Source · Tools · Benchmarks

· unitree_lerobot: Unitree's released training stack for the G1 dual-arm dexterous-hand humanoid, adapted from the LeRobot framework and connecting hardware — embodiment data — imitation learning pipelines, with about 708 GitHub stars source (WeChat, CN)

· GraspXL: an open cross-hand-morphology grasping dataset with grasp motion sequences for 500,000+ objects, covering MANO, Allegro, Shadow and other hand morphologies, for training morphology-adaptation and contact priors source (WeChat, CN)

· UniDex (CVPR 2026): learns general-purpose dexterous-hand control from egocentric human videos, together with EPIC-Contact confirming that "egocentric human video" is becoming a core data asset for dexterous-hand policies source (WeChat, CN)

· World model benchmarks: a wave of new benchmarks — WorldModelBench, WorldScore, WorldLens, and others — has launched in quick succession, responding to the confusion of video generation, embodied AI, and autonomous driving each claiming to build "world models" without a unified way to measure them source (WeChat, CN)

II. Financing & Deals

Hidriv (希迪智驾) ｜ Hong Kong Stock Exchange Main Board filing ｜ CICC, CSC, Ping An as joint sponsors · autonomy

Hidriv, focused on autonomous driving in enclosed environments, has again filed for listing on the Hong Kong Stock Exchange, specializing in autonomous mining and logistics vehicles and V2X vehicle-road coordination. As of mid-2025, it had cumulatively delivered 304 autonomous mining trucks and 110 standalone truck systems, plus indicative orders for another 357 mining trucks; by 2024 product sales revenue, it ranks first in China's autonomous mining-truck solutions market, and has deployed the world's largest mixed fleet of autonomous and human-driven mining vehicles. Financially it remains in an investment phase — 2024 revenue was RMB 410 million with a 24.7% gross margin, but the company posted a loss of RMB 581 million that year.Source: Sina Finance source

Vita Dynamics (Vbot) ｜ Pre-A round ｜ with participation from Yuanjing Capital · embodied

Quadruped robot company Vita Dynamics completed a Pre-A financing round with participation from Yuanjing Capital; the company disclosed that its "Super Robot Dog" has begun mass-production delivery, moving from demo units to actual shipments — a rare mass-production milestone in the quadruped-platform race.Source: Yuanjing family source

UBTECH ｜ CSRC filing for full circulation of H shares · humanoid

UBTECH, the Hong Kong-listed humanoid robot company, has received a filing notice from the China Securities Regulatory Commission (CSRC) regarding full circulation of its H shares, which will allow domestic shareholders' shares to convert into tradable H shares — further opening up its financing and liquidity channels in Hong Kong.Source: Ifeng source

Market ｜ Embodied AI financing in China reached roughly RMB 93.5 billion in H1 · humanoid ⚠️ statistical methodology varies

Multiple institutions reviewing H1 activity report that financing in China's embodied AI sector totaled roughly RMB 93.5 billion, with about 322 deals in June alone; full-stack humanoid companies drew the heaviest capital. This extends a recent fundraising wave in which both AgiBot (Zhiyuan) and Zhiping Fang (智平方) reached RMB 20 billion valuations. However, revenue and order books at these funded companies generally fall far short of supporting valuations in the tens of billions of RMB — "buying a ticket to the future" remains the honest description of most of the capital deployed right now.Source: FutureIndustryMax source (WeChat, CN)

III. Commercialization & Deployment

CATL puts Galbot's heavy-payload humanoid S1 to work on its own battery lines · industrial

CATL and embodied-AI company Galbot have signed a global strategic partnership, with the core being that Galbot's heavy-payload humanoid robot Galbot S1 is now operating routinely on CATL's smart production lines. According to both companies, the S1 has a dual-arm payload of 50kg, centimeter-level vision-only localization, and 360° omnidirectional obstacle avoidance; paired with CATL battery cells (claimed to achieve a cell failure rate at the parts-per-billion level), it can run continuously for about 8 hours per charge and has been deployed for module and battery-pack manufacturing, directly replacing high-intensity tasks such as material handling and pick-and-place. The partnership also extends CATL's "Ningjia Service" network — which covers inspection, maintenance, and recycling for power batteries — to humanoid robots, with both companies saying they will jointly build the world's first after-sales service standard for humanoid robots. Following earlier cases in 3C quality inspection and automotive production lines, this is another example of a manufacturing leader using its own use cases to drive humanoid adoption while locking in core component supply — this time backing it with batteries, deployment scenarios, and after-sales service all at once.Source: Markets Insider source

WeRide × Grab: Singapore's first public robotaxi operation · autonomy

WeRide and Southeast Asian super-app Grab have officially launched public operations of the Ai.R autonomous mobility service in Punggol, Singapore — the first autonomous public mobility service there open to the general public. Following a three-month trial with over 1,000 residents participating, the fleet has logged more than 30,000 km safely; the service is now available for free public booking and is set to transition to commercialization by mid-2026. The vehicles are the Robotaxi GXR model, which is already in routine operation in China, the UAE, and Saudi Arabia. During the initial operating period, each vehicle still carries a safety operator, and many of these safety operators have transitioned from being experienced Grab drivers through joint training programs (14 have already been certified) — a model combining "technology deployment with driver retraining."Source: Autohome source

Pony.ai connects to a local ride-hailing platform in Singapore; claims single-city profitability in Guangzhou and Shenzhen · autonomy ⚠️ single-city profitability is company-reported

Racing WeRide in the same city, Pony.ai announced that its autonomous mobility service in Singapore has officially connected to the local ride-hailing platform Zig and opened to public bookings, and claimed it has achieved single-city profitability in both Guangzhou and Shenzhen. Around the same time, WeRide disclosed Q1 2026 revenue of RMB 114 million, up 58% year-on-year, with net loss narrowing about 63% year-on-year — figures from both companies pointing to robotaxis moving from technology validation toward operational payoff.Source: Netease Auto (Wtoutiao) source

IV. Industry Developments

UN global technical regulation on autonomous driving approved, ushering in a unified standards era for L4 · autonomy

The UN Global Technical Regulation on Automated Driving Systems, jointly led by China, the EU, the US, the UK, Canada, and Japan, has been formally approved and published following a vote by all contracting parties, providing a unified technical standards framework for compliant deployment of L4 autonomous driving worldwide — seen as a policy inflection point for robotaxis moving from regional pilots to global operations. The cost curve underpinning commercialization is also falling fast: bulk-purchase prices for a 128-line LiDAR unit have dropped more than 90% compared with a decade ago, and Baidu's Apollo Go sixth-generation mass-produced autonomous vehicle has fallen to a unit price of RMB 204,600. Separately, in a comprehensive robotaxi ranking published by third-party firm AutonomyAI, Apollo Go ranked first, Waymo second, and Tesla fifth (per that ranking's methodology, for reference only). The regulation taking effect is not an endpoint but the starting point for the industry shifting from a "technology narrative" to an "operations narrative" — the real test lies in long-tail road conditions, safety systems, and public trust.Source: Netease Auto (Wtoutiao) source

China's robot exports approached RMB 20 billion in the first five months, with cleaning robots accounting for 70% · industrial

The latest customs data shows that in the first five months of this year, China exported a total of 10.377 million robots of various kinds, with total export value of RMB 19.99 billion; the products have reached more than 150 countries and regions worldwide, with the EU and ASEAN as the primary destinations. Of this, cleaning robots alone accounted for more than 70% at RMB 14 billion, industrial robot exports totaled about 70,000 units, and exports of intelligent bionic robots exceeded 8,000 units for the first time — the export mix shows that the globalization of Chinese robots is still led mainly by mature consumer/industrial categories, with bionic/humanoid robots still at an early ramp-up stage.Source: Sina Finance source

Matrix Robotics claims Q3 mass-production delivery of MATRIX-3, targeting about 1,000 units for the year · humanoid ⚠️ company-reported

Matrix Robotics, founded by Zhang Haixing (former head of Tesla's China design research center), says its flagship humanoid MATRIX-3 will enter mass production and delivery in Q3 this year, targeting about 1,000 units for the full year. It uses the same "linear joint" technical route as Tesla's Optimus (33 whole-body degrees of freedom, 27-DoF dexterous hands), with the standard version starting at RMB 580,000. The founder claimed to have "made good on what Musk only talked about." This is a production plan and capability claim; delivery, yield, and real-world industrial performance remain to be proven, so its timeline should be viewed skeptically.Source: Jiemian News source

Humanoid company's KinetIQ Ascend: real-world reinforcement learning on a production VLA · embodied ⚠️ company-reported

Robotics company Humanoid unveiled KinetIQ Ascend, which it describes as the first publicly disclosed case of running end-to-end vision-based reinforcement learning on a production-grade VLA on a real dual-arm humanoid robot — with the robot practicing on its own around real production-line tasks. Per its disclosure, throughput for a goods sorting/handover task rose 85% with success rate climbing from 80% to 98%, dual-arm carrying throughput doubled with success rate rising from 78% to 99%, and speed reached 1.5x that of the original human demonstrations. The figures are striking but are all self-reported production-line numbers from the company, still lacking independent replication.Source: The Robot Report source

Fei-Fei Li's essay proposes a taxonomy for "world models": renderers / planners / simulators · world-model

With video generation, robotics, and NVIDIA all claiming to build "world models," Fei-Fei Li published a long essay proposing a functional taxonomy: based on "what the model is actually doing," she divides them into renderers that merely aim for good-looking pixels, planners that output correct actions for an agent, and simulators that model states and their consequences. This is a conceptual framework and viewpoint rather than a new result, but it offers a coordinate system for clarifying use cases and enabling apples-to-apples comparison amid the current rush to claim "world model" status.Source: ChainCatcher source

Hardware · Supply Chain

· Tactile sensors: PaXini Perception released the all-solid-state array-type PX-Cloud V3, reaching 32×32 (1,024 cells) at under RMB 2,000 per unit, compared with a comparable 16×16 Tekscan unit priced at RMB 12,000. After integration into AgiBot's Lingxi Pro, success rate across 12 fine-manipulation tasks reached 95.3% (up 42 percentage points versus vision-guidance alone, per company testing); Unitree's H1 second generation has also integrated a Chinese-made 32×32 array. Industry estimates put the tactile-sensing cost per humanoid unit dropping from about RMB 200,000 to under RMB 30,000 source (WeChat, CN)

· Joint modules: Integrated-joint-module maker Quanzhibo, after completing an A+++ round (led by Hillhouse Venture Capital, with strategic investment from AgiBot and Linker Bionics), disclosed monthly shipments exceeding 60,000 units, reflecting how humanoid ramp-up is driving firm demand for upstream modules source (WeChat, CN)

· Dexterous-hand transmission: A tendon-driven dexterous-hand startup founded by a post-95s entrepreneur has raised nearly RMB 1 billion in cumulative financing over a year and a half, with Li Auto and Xiaomi both racing to invest; it specializes in a mass-producible pure tendon-drive design, spec'd at 25 degrees of freedom / 380 grams / 30kg+ payload / 0.6-second open-close cycle / 0.2mm positioning accuracy (per company specifications) source (WeChat, CN)

FutureX · Physical AI Daily — Issue 48 (07/05)

Shawn — Sat, 04 Jul 2026 14:49:31 +0000

Today's Highlights

· The China Humanoid Robot 100 Council and the China Machinery Industry Federation jointly issued an initiative to regulate the development and marketing of "emotional companion" humanoid robots, urging strict ethical and privacy safeguards — widely read in the market as aimed squarely at models like UBTech's U1, a bionic companion robot priced near RMB 1 million.

· Tesla brought unsupervised Robotaxi service to Texas for the first time, launching in Miami (Model Y), expanding its operating footprint to 5 cities as Q2 deliveries rose against the trend by roughly 25%.

· Fei-Fei Li and NVIDIA GEAR's SimFoundry: a single real-world video is enough to automatically generate an interactive simulation plus "digital cousins," boosting real-robot task success rates by up to 40% and reaching a 0.911 correlation between sim evaluation and real-robot performance.

· LeCun's team's AdaJEPA lets world models "learn on the job" via test-time adaptation, nearly doubling out-of-distribution planning success rates while adding only 0.01–0.03 seconds of extra latency.

· Guangxiang Technology (Tsinghua-incubated startup), spun out of Tsinghua, closed an angel round of hundreds of millions of yuan, championing a "physically-native foundation model" approach; its industrial robots have already run 21.5 hours straight on an auto production line with zero errors.

I. Research Progress

SimFoundry: Automatically Generating Trainable, Evaluable Simulated Worlds from a Single Real Video · benchmark

Collecting real-robot data is expensive and hard to scale, and Real2Sim approaches try to build simulations backward from real video — but past solutions often solved only one piece of the puzzle, either reconstructing 3D scenes or enabling policy evaluation alone. SimFoundry chains together scene reconstruction, data generation, policy evaluation, and policy training into a complete real-to-sim-to-real pipeline: an ordinary RGB video automatically generates an interactive "digital twin," from which "digital cousins" (with swapped objects, altered layouts, or new tasks) are batch-derived while preserving object function and affordances, yielding near-limitless training data. After introducing three categories of digital cousins, real-robot task success rates rose by 17%, 21%, and 40% respectively, sim-to-real Pearson correlation reached 0.911, and policies trained solely on this generated data could deploy zero-shot to real robots.

Nadun Ranawaka Arachchige et al. (NVIDIA GEAR / Georgia Tech / Stanford / UT Austin / University of Toronto, including Fei-Fei Li, Jim Fan, and Yuke Zhu) · arXiv 2606.28276 source · Coverage: QbitAI source (WeChat, CN)

AdaJEPA: Letting World Models Keep Learning After Deployment · world-model

Past JEPA-family latent-space world models have generally frozen their parameters after training, which makes them prone to failure when the test-time distribution shifts, while short-horizon MPC rollouts can further amplify single-step errors. NYU and LeCun's startup AMI propose AdaJEPA, bringing test-time adaptation (TTA) into world models: after executing only the first segment of an MPC-planned action, the next real observed frame is used as a self-supervised signal to lightly update the final layers of the encoder and predictor (just one gradient step per timestep). On out-of-distribution tasks like PushObj and PointMaze, planning success rates nearly doubled (PointMaze GD rose from 53.3% to 78.7%), while added latency was only 0.01–0.03 seconds — more like giving a frozen model a "self-calibration at deployment" mechanism.

Ying Wang et al. (NYU CILVR Lab / AMI, advised by Mengye Ren and Yann LeCun) · arXiv 2606.32026 source · Coverage: QbitAI source (WeChat, CN)

EgoTSR: Teaching VLMs to Judge Whether a Task Is Actually Progressing · perception

An arm fumbles a grasp and the cup falls back onto the table — the later frame in the sequence, yet the task has actually reverted to its starting point — and VLMs often get this backwards, relying on the shortcut that "a later-appearing frame looks more complete." EgoTSR, from a five-university team including Zhejiang University, targets exactly this "temporal order bias": feeding the model both the forward and reversed order of the same image pair exposes the shortcut (InternVL-8B scores nearly 99% accuracy forward but collapses to roughly 2% in reverse). The team built a 46-million-sample dataset and trained with a three-stage "explain first, then internalize, then plan" curriculum, adding a subtask planner for long-horizon progress reasoning; the result is 92.4% average accuracy on long-horizon tasks with only a 0.1-percentage-point gap between forward and reverse order, and it can output a continuous task-completion curve for long-video monitoring.

Zhejiang University / Tianjin University / Shanghai Jiao Tong University / National University of Singapore et al. (ICML 2026) · arXiv 2604.10517 source · Code: Collab-Gen/EgoTSR · Coverage: Jiqizhixin (Synced) source (WeChat, CN)

GEM: Giving VLA "Depth Vision" to Fill the Spatial Gap · vla

Most current VLAs are "semantic giants, spatial dwarfs" — they can recognize a "red cup" but can't accurately judge how far away it is. Tencent's team behind GEM brings depth-map generation into the VLM pretraining stage: a diffusion-based depth generation head is attached to the side of a Qwen3-VL backbone, conditioned on visual tokens, forcing the visual representation to encode 3D structure. On the VSI-Bench spatial understanding benchmark, scores rose from 57.9 to 70.6 (+12.7%), surpassing Gemini-3-Pro, with an average of 96.1% across four LIBERO suites; on real-world UR5 long-horizon tasks (Table Bussing), average progress improved 67% over π₀.₅, and removing depth supervision caused a significant performance drop, confirming it functions as core infrastructure rather than an add-on.

Tencent · arXiv 2605.28548 source · Coverage: Embodied AI Manufacturing source (WeChat, CN)

Drop-Then-Recovery: Cutting Half the Language Layers Makes VLA Stronger · vla

Is a multi-billion-parameter "language brain" really necessary for VLA? University of Maryland and Cisco Research's Drop-Then-Recovery offers a counterintuitive answer: physically removing half the language backbone and then fine-tuning to recover actually raised success rates rather than lowering them — OpenVLA-OFT on LIBERO went from 95.0% to 98.3%, and π0.5 went from 91.7% to 94.0%; meanwhile the vision and action pathways collapsed after removing even a tiny fraction of parameters, revealing a clear asymmetry of "redundant language, untouchable action." The authors propose a GateProbe virtual-gate metric to predict "which layers can be safely removed and recovered," and the finding also serves as a warning that current manipulation benchmarks' tests of language grounding may be too weak.

University of Maryland / Cisco Research · Coverage: Paper Digest Hall source (WeChat, CN)

OmniContact: Linking Up Long-Horizon Humanoid Manipulation with "Contact Flow" · locomotion

The hard part of long-horizon humanoid loco-manipulation is often the "seams between actions": if a box shifts, a suitcase gets stuck, or the previous segment wasn't executed precisely, can the next segment still pick up where it left off? Noitom Robotics and HKUST et al. propose OmniContact, which uses "contact flow" (who touches what, when, and how the body moves before and after contact) as a sparse intermediate interface: a high-level CF-Gen generates the contact flow, while a low-level CF-Track uses reinforcement learning to track it into full-body motion. With online replanning added, box-carrying correction success reached 99.7%, box-pushing rose to 94.5%, and the system can connect to a VLM to break down semantic tasks (such as arranging scattered boxes into a heart shape) into object-level goals.

Noitom Robotics / Hong Kong University of Science and Technology / Wuhan University / University of Hong Kong · Project page: omnicontact.github.io · Coverage: Embodied AI Research Lab source (WeChat, CN)

Other papers today: world models remain the day's most active research direction — WM-AMT injects "predict-future-states" capability before post-training, letting agents perform what-if reasoning first (claiming roughly +9.8% reasoning accuracy); LoopWM lets a world model "reconsider" the same step repeatedly before deciding; and there's also an efficient reinforcement learning framework that connects a world model to a real robot to mitigate "dream-like" visual hallucinations and contact-dynamics artifacts.

Open Source · Tools · Benchmarks

· Xiaomi open-sources its latest autonomous driving model: led by its core technical team, focused on handling complex driving scenarios with a small model source

II. Funding & Deals

Guangxiang Technology (Tsinghua-incubated startup) ｜ Angel round (cumulative) ｜ Hundreds of millions of yuan · world-model

This round included financial and industrial investors such as Zhuhai Science & Technology Industry Group, Xingzheng Capital, Songhe Capital, Shunxi Fund, and SeeFund, along with listed company Xingyun Technology, plus additional investment from existing shareholders Nova Capital and L2F Guangyuan Fund. Guangxiang Technology was founded in April 2025, incubated jointly by Tsinghua University's School of Vehicle and Mobility and School of Artificial Intelligence. Its CEO, Zhang Tao, is the former head of Amap's spatial perception engine, and co-founder Li Shengbo is an expert in autonomous driving reinforcement learning. Technically, the company has chosen a "physically-native foundation model" path distinct from mainstream VLA and video-prediction world models — letting the model spontaneously develop an understanding of mass, friction, and causality through physical interaction, supported by its Phi-RL Matrix algorithm, Phi-Space data assets, and Phi-Arch platform. Its industrial robot, Phi-Bot X1, has already completed 21.5 continuous hours of welding pickup and placement on an auto production line with zero errors, improved mobile quality-inspection efficiency by 25–45% over manual work, and formed partnerships with several leading automakers, targeting the "30% digitalization gap" that robotic arms and PLCs can't handle.

Source: Hard Krypton source (WeChat, CN), 36Kr source

Huiguang Innovation ｜ Seed + angel round ｜ Tens of millions of yuan · hardware

Robot tactile sensor and tactile-data solution provider Huiguang Innovation (a team of Tsinghua graduates born after 2000) has closed two consecutive funding rounds, led by Vertex Ventures and a dual-currency financial investor, with follow-on participation from Poqoo Robotics, Infinity Fund SEE Fund, and the Shuimu Tsinghua Alumni Seed Fund. Tactile sensing is a key bottleneck for dexterous manipulation and force control, and this round continues capital's bet on "embodied perception layer" components. Source: The Beauty of Algorithms and Mathematics source (WeChat, CN)

Ouster ｜ Follow-on offering ｜ Roughly $200 million · hardware

Lidar maker Ouster raised roughly $200 million to bolster cash reserves and scale up production, adding fresh capacity to the robotics/autonomous-driving perception supply chain. Source: Pluang source

RoboParty ｜ Pre-A round ｜ Amount undisclosed · embodied

RoboParty closed a Pre-A funding round, with investment from battery giant CATL — another signal of industrial capital extending along the "battery-to-robotics" chain into embodied AI. Source: Sina Finance source

Overland AI ｜ Military contract ｜ Nearly $20 million · autonomy

Autonomous driving company Overland AI won a roughly $19.7 million contract from the U.S. Marine Corps to build unmanned autonomous military vehicles — a defense-sector deal offering a practical monetization path for off-road autonomous driving. Source: GeekWire source

Xiaoyu Zhizao (Xiaomi-affiliated) ｜ Series B+ ｜ Hundreds of millions of yuan · industrial ⚠️ Single-source claim

Industrial embodied AI platform Xiaoyu Zhizao reportedly closed a Series B+ round of hundreds of millions of yuan, pursuing a "platform" approach in an attempt to sidestep the crowded competition in full humanoid systems. Details on the exact amount and round terms remain unconfirmed by authoritative sources and are based on the company's own account. Source: Selected Business Plans source (WeChat, CN)

III. Commercialization & Deployment

Tesla's Unsupervised Robotaxi Arrives in Miami, Operating Footprint Grows to 5 Cities · autonomy

Tesla brought Robotaxi to Texas for the first time, launching in Miami using the Model Y and describing the service as "fully unsupervised." Its operating area has now expanded to roughly 5 cities: Miami, Dallas, Houston (unsupervised), and Austin (hybrid mode, with safety monitors in some vehicles), putting it in direct competition with Waymo. The Miami launch coincided with Tesla's disclosure that Q2 deliveries rose against the trend by roughly 25%, giving a short-term boost to market confidence in its autonomous-driving narrative. That said, "unsupervised" operation is still bound by geofencing (a restricted operating area) and remote support, and the scale and real-world safety performance still require longer-term data to verify. Source: blockchain.news source, Refresh Miami source

Waymo Opens Fully Driverless Rides in Nashville · autonomy

Waymo officially opened fully autonomous ride-hailing service in Nashville, continuing the steady expansion of its Robotaxi network across U.S. cities. Source: Mashable source

XPeng's First GX Robotaxi Production Unit Rolls Off the Line · autonomy

XPeng's GX Robotaxi has produced its first mass-production unit, marking a milestone from prototype to production; separately, XPeng said its VLA 2.0 will achieve "map-free" autonomous driving for international markets by 2027. Source: autohome.com.cn source

GM Cuts 1,000 Jobs at Detroit EV Plant, Deploys 50 AI Collaborative Robots · industrial

General Motors cut roughly 1,000 jobs at its Detroit EV hub while deploying 50 AI collaborative robots, drawing strong pushback from the union. This is a stark example of the "machines replacing workers" tension arising as embodied automation advances, and a reminder that deployment pace is being pulled back by employment and social pressures. Source: finance.biggo.com source, Futurism source

Household Embodied Robots Start Entering Homes via "Pay-Per-Use" Rental · humanoid

Household embodied robots are exploring the consumer market through "pay-per-use" rental, with one service offering chore trials for around RMB 74 per 3 hours. Zhiyuan Robotics (Chinese humanoid startup) has also registered the trademark "RoboShare" for its open rental platform "Qingtian Zu" ("Rent-a-Sky"), attempting to turn humanoid robots into a rentable service. The real scale of usage, reliability, and repeat-rental rates for this model remain to be verified. Source: Pandaily source, Trademark Supermarket source (WeChat, CN)

IV. Industry Developments

Two Major Industry Bodies Jointly Issue Guidelines on "Emotional Companion" Humanoid Robots · humanoid

The China Humanoid Robot 100 Council and the China Machinery Industry Federation jointly issued an initiative stating that the industry should preserve humans' prerogative to set preferences for service, follow safety and ethical standards in product design and advertising, strengthen personal information protection, and reinforce quality management to guard against privacy breaches and physical harm to users, while also calling for breakthroughs in core technologies and real-world application. This marks the first time an authoritative Chinese industry body has set explicit ethical and safety boundaries for "companion/emotional" humanoid robots, timed just as UBTech's U1 (top configuration priced at RMB 990,000 for the male version and RMB 880,000 for the female version, marketed for long-term emotional companionship) saw pre-orders exceed 13,000 units, with its appearance and positioning continuing to draw controversy. It's worth noting that this is industry-level guidance and self-regulation, not a regulatory ban; some secondary-market commentary framing it as "invalidating the commercialization logic" overreaches, but it does draw compliance boundaries around this most imaginative — and most sensitive — segment of the industry. Source: China News Service source

UN Approves Global Autonomous Driving Regulations; XPeng Plans Map-Free Overseas Rollout by 2027 · autonomy

The United Nations has approved global regulations related to autonomous driving, paving the way for cross-border deployment of advanced autonomous driving. XPeng used the occasion to state that its VLA 2.0 will achieve "map-free" autonomous driving for overseas/international markets by 2027; CEO He Xiaopeng separately said L4-to-L5 autonomy could arrive within 3 to 5 years. The regulatory framework progress is factual, but the specific timing and capability of mass production and overseas rollout remain forward-looking plans from the manufacturer. Source: netscn.com (Wangtongshe) source

Unitree IPO Aftermath Continues; Founder Chen Li Discusses the "ChatGPT Moment for Embodied AI" · humanoid ⚠️ Company statement

Following the approval of its IPO registration on the STAR Market (previously reported), the A-share robotics sector rally driven by Unitree continued today, and Unitree Robotics was selected for an on-site regulatory inspection. Founder Chen Li publicly stated that embodied AI's "ChatGPT moment" requires reaching "two 80% thresholds," and predicted the company will lead a new wave of consumer products by 2030 — a stated vision, not an established fact. Source: cls.cn (China Securities Journal/CAIXIN) source, Sina Finance source

12 New Occupations, Including "Embodied AI Robot Application Technician," Announced for Public Comment · adjacent

12 new occupations were announced for public comment, including "Embodied AI Robot Application Technician" — a recognition at the occupational-classification level that reflects the industry's shift from technology validation toward large-scale employment and standardized job roles. Source: Sina Finance source

Shanghai International Embodied AI Expo Closes with Over 120 Partnership Agreements · adjacent

China's first independent professional exhibition dedicated to "embodied AI," CIEI 2026, has closed, with nearly 200 companies — including Unitree, Zhiyuan Robotics, UBTech, LimX Dynamics, Galbot, and Fourier Intelligence — exhibiting everything from full robotic systems to core components, and reaching more than 120 partnership agreements on-site. Source: MiaoZhuo Talks AI source (WeChat, CN)

Hardware · Supply Chain

· Dexterous Hand Multimodal Dedicated Chip: The Zhongke group (Zhongke Alpha / Zhongke Semiconductor) has launched a multimodal sensing and control chip dedicated to next-generation dexterous hands, presenting three core chip solutions using a 22-DOF dexterous hand as an example, emphasizing high precision, low latency, and small form factor source (WeChat, CN)

· Yuequan Bionics Y-Hand M1: This new dexterous hand returns to "human hand biology," balancing rigidity and compliance to combine grip strength with gentle, damage-free handling of fragile or irregularly shaped objects source (WeChat, CN)

· Li Auto Mach M100 Chip: Li Auto's self-developed autonomous-driving chip, the Mach M100, has debuted, marking further progress in in-house vehicle compute development source

· Harmonic Reducer Supply Gap: A brokerage research report claims humanoid robot harmonic reducers face a roughly 20% supply gap, forming the core thesis behind upstream capacity-flexibility investment plays (⚠️ analyst report claim, not measured data) source (WeChat, CN)

FutureX · Physical AI Daily — Issue 47 (07/04)

Shawn — Fri, 03 Jul 2026 14:44:08 +0000

Today's Highlights

· South Korean conglomerates Samsung, Hyundai, SK and others announced a combined roughly ₩312 trillion (about $195 billion) in southern-region industrial investments on the same day, betting heavily on physical AI manufacturing — Samsung alone will invest ₩60 trillion to build a physical AI manufacturing cluster in the Yeongnam region, plus another ₩19 trillion for a robotics factory in Gumi.

· Industrial robot leader Estun Automation (Chinese industrial robotics maker) plans to fully acquire its collaborative-robot subsidiary Estun Cobot to expand into embodied AI, with the stock hitting the daily limit-up 3 times in 4 days.

· XBOT, a food-service embodied AI company founded by former Xiaomi executive Tang Mu, closed two funding rounds totaling hundreds of millions of yuan (Series A: ¥200 million; Series B: ¥300–500 million).

· Unitree Robotics' (Chinese humanoid robot maker) IPO registration approval ignited China's A-share robotics sector, with over 40 stocks hitting limit-up in a single day (the IPO was approved yesterday; today's move reflects secondary-market trading).

· Autonomous driving company Momenta (Chinese self-driving startup) saw its Hong Kong IPO public offering oversubscribed by more than 414 times, as the Hong Kong listing window for robotics/autonomous-driving companies stays hot.

1. Research Progress

Learn to Move Before You Learn to Act: Task-Agnostic Pretraining for VLA (TAP) · vla

VLA models have long been bottlenecked by scarce expert demonstration data. This paper separates "learning how to move" (motor competence) from "learning what to do" (semantic alignment): it first uses cheap, unlabeled data — including discarded off-task trajectories and autonomous robot play — to learn transferable motion priors via self-supervised inverse dynamics, then aligns those priors to language instructions using very few labeled demonstrations, significantly reducing reliance on expensive demos. It gained HF↑3 on the day, and the approach has broad value for data-constrained settings.

Junhao Shi et al. · arXiv 2607.02466 source

One Demo Is Enough: Real-World Robot Reinforcement Learning (AutoSERL) · manipulation

Real-world RL faces two major pain points: expensive data and the need for continuous human intervention during training. AutoSERL fully automates the intervention process using just a single demonstration, achieving 100% success rate on insertion tasks and greater robustness to positional perturbations — cutting the startup cost of real-world RL down to one demo.

Yuwan Liu et al. · arXiv 2607.01651 source

PhysMani: A Physics-Grounded 3D World Model for Dynamic Manipulation · world-model

Grasping fast-moving targets remains a hard problem for embodied AI, and existing VLA and world models both struggle to jointly achieve precise 3D geometry and physically plausible predictions. PhysMani couples a physics-grounded 3D Gaussian world model with a "future-aware" action policy, so predictions respect both geometry and physics, for dynamic object manipulation in unstructured 3D environments.

Peng Yun et al. · arXiv 2607.01938 source

Bridge-WA: Predicting Only Where and How the World Changes, Not Rendering Full Future Frames · world-model

General-purpose VLA benefits from large vision-language priors, but effective manipulation also requires anticipating action-relevant scene changes; large generative world models that do dense future rollouts are expensive, spending much of their compute on visual details only weakly related to control. Bridge-WA distills a frozen "future-change teacher" into three compact priors (future tokens, change maps, motion-flow maps), lightly boosting task success and progress — a pragmatic route of "predicting less pixels, more causality."

Yongjie Bai et al. · arXiv 2607.02195 source

VT-WAM: A Visual-Tactile World-Action Model for Contact-Rich Manipulation · world-model

Contact-rich manipulation requires instant response to local deformation, pressure, slip, and friction — cues that are sparse or even invisible in vision. Existing visual-tactile policies mostly feed tactile signals directly into action prediction, rarely modeling the dynamics of tactile deformation during action generation. VT-WAM brings tactile sensing into the world model's predictive loop, filling this gap.

Shuai Tian et al. · arXiv 2607.02503 source

Imagining Touch: "Tactile-Informed" Manipulation Without Tactile Hardware · perception

Tactile sensing can substantially improve contact-rich manipulation, but sensors are fragile, need calibration, and are costly to maintain, limiting real-world deployment. This paper proposes imagined tactile representations — at deployment, no tactile sensor is installed, yet the robot still gains the benefits of tactile knowledge, answering the fundamental question of whether tactile benefits are attainable without tactile hardware.

Zhiyuan Zhang et al. · arXiv 2607.01684 source

Guided Action Flow: Adding Inference-Time Q-Guidance to Flow-Matching VLA · vla

Flow-matching VLA generates action chunks through iterative transport, naturally leaving room for "no retraining needed" test-time guidance. This framework keeps a pretrained SmolVLA policy frozen and uses a learned action-chunk critic to guide its reverse flow sampler, improving performance without touching the backbone weights — another example of the low-cost "frozen backbone + inference-time polishing" paradigm.

Liuhaichen Yang et al. · arXiv 2607.02092 source

HEFT: Heavy-Payload Teleoperation for Full-Scale Humanoids · locomotion

General motion tracking/teleoperation is one path to scaling humanoid skills, but most frameworks are validated on small platforms or without real payloads, leaving full-scale humanoids under real load almost entirely unstudied. HEFT learns from noisy VR references using privileged motion guidance, then applies a windowed payload curriculum to progressively add load, achieving robust heavy-payload tracking.

Chenxin Liu et al. · arXiv 2607.02332 source

Other papers today: VLA-Corrector adds lightweight "detect-and-correct" reasoning with on-demand adaptive action horizons to chunked VLA (arXiv 2607.01804 source); The Moving Eye uses a dual-arm setup where one arm operates while the other serves as a moving camera, improving VLA spatial generalization (arXiv 2607.02322 source); WorldSample closes the "real-synthetic" data augmentation loop in real-world RL using a world model (arXiv 2607.02431 source); ACID uses inverse-dynamics consistency to constrain the achievability of intermediate states in world-model planning (arXiv 2607.02403 source); Actuator Reality Shaping shapes actuator dynamics for zero-shot sim-to-real transfer (arXiv 2607.02205 source); VLAFlow proposes a unified training framework for cross-comparing different robot pretraining paradigms (arXiv 2607.01586 source); Neuro-Symbolic Safety Guidance uses constrained flow matching to give VLA predictive obstacle avoidance (arXiv 2607.01378 source); Controllable Sim Agents generates traffic simulation agents controllable along interpretable axes via behavioral latent variables (arXiv 2607.02496 source).

Open Source · Tools · Benchmarks

· Embodied.cpp: A portable embodied-model inference runtime for heterogeneous robots, unifying deployment of VLA and world-action models (WAM), providing modular multi-rate execution, latency-first fused inference, and extensible operators/IO — easing the fragmentation of "each model with its own Python stack plus robot-side glue code" (arXiv 2607.02501 source).

· CommonRoad-Game: A lightweight human-in-the-loop autonomous driving simulation framework, tightly coupled with the CommonRoad platform, specifically built for systematically testing motion planners in interactive scenarios involving humans and analyzing human driving behavior (arXiv 2607.01382 source).

· DL-VINS-Factory: A modular framework unifying learned visual front-ends (ALIKED, SuperPoint, XFeat, etc.) with LK optical flow or LightGlue matching for visual-inertial SLAM, enabling systematic evaluation of the practical value of deep features in tightly-coupled VI-SLAM (arXiv 2607.01757 source).

2. Funding & Deals

XBOT | Series A + Series B | Hundreds of Millions of Yuan Combined · embodied

General-purpose food-service embodied robotics company XBOT closed two consecutive funding rounds: a ¥200 million Series A funded by Hong Kong's Jiankun Capital (GPTX), and a ¥300–500 million Series B with participation from multiple government funds, USD funds, and industry partners. The company was founded by former Xiaomi executive Tang Mu, and on the same day launched an in-house coffee brand to validate embodied commercialization through "a single cup of coffee." Funds will go toward R&D, market expansion, and team building.Source: Tech Capital Circle source (WeChat, CN), Robotics Outlook source (WeChat, CN)

Quanzhibo | Series A+++ | Led by GL Ventures · hardware

Wuxi-based Quanzhibo (a maker of integrated robotic joint modules) closed a Series A+++ round led by GL Ventures (Hillhouse's venture arm), with Zhiyuan Robotics (Chinese humanoid robot startup) and Lingxin Qiaoshou (Chinese dexterous-hand maker) joining as strategic industry investors. Both robot makers and top-tier capital are now betting directly on upstream joint modules, extending this cycle's trend of "capital flowing from full robots toward components like joint modules and dexterous hands."Source: PEDaily source, Shouchuang Holdings source (WeChat, CN)

Rushen Robotics | Pre-A Round | ¥100 Million · embodied

Shanghai-based Rushen Robotics (founded by a Tsinghua professor, focused on elderly-care scenarios for embodied AI) closed a ¥100 million Pre-A round, with investment from Qingsong Capital, Runze Technology, and Pinghu Zexin, to accelerate deployment of embodied AI in elderly-care facilities and home settings.Source: Robotics Outlook source (WeChat, CN)

Momenta (Shanghai Autonomous Driving Technology) | Hong Kong IPO | 414x Oversubscribed · autonomy

Autonomous driving solution provider Momenta's Hong Kong IPO public offering was oversubscribed roughly 414 times, adding another hot listing to the recent wave of Hong Kong IPOs. Investor enthusiasm for advanced autonomous-driving and embodied-AI names mirrors the rally in China's A-share robotics sector.Source: South China Morning Post source

Lishang LISSOME | Series A | Tens of Millions of Yuan · embodied

AI kitchen robot brand Lishang closed a tens-of-millions-of-yuan Series A round led by Sequoia China and Brizan Ventures, targeting embodied AI deployment in the high-frequency home cooking scenario.Source: Embodied Universe source (WeChat, CN)

3. Commercialization & Deployment

Kodiak Completes Autonomous Trucking Program in Ohio · autonomy

Kodiak announced the completion of its autonomous trucking program in Ohio, marking another real-world milestone for driverless long-haul freight; the autonomous trucking sector has recently been making steady progress on "mileage/program delivery," not just demos.Source: Investing.com source

Tesla Robotaxi Expands to Florida, Its Third Operating State · autonomy

Tesla carved out a small Robotaxi operating zone in Miami, making Florida its third state for autonomous-driving deployment. But media reports also note that Tesla is still struggling to scale up in Texas — a clear gap remains between expanding to new states and actually scaling capacity.Source: Electrek source

Waymo Takes First Step into Southern Europe · autonomy

Waymo has taken its first step into southern Europe, continuing to accelerate its overseas/international expansion. Following adjustments to its Phoenix partnership, Waymo's focus is increasingly shifting toward overseas/international cities and new partnerships.Source: eletric-vehicles.com source

UBTech's "Kissonic" Industrial Humanoid Debuts With 4,000 Pre-Orders · humanoid

UBTech (Chinese humanoid robotics maker) launched its industrial-native humanoid robot "Kissonic," accompanied by an industrial embodied model called "FabriX." The company claims it secured 4,000 units in orders at launch. The figure is a vendor-disclosed order count; actual installations and delivery pace remain to be verified.⚠️ Vendor-disclosed figuresSource: Xingfu Feidong source (WeChat, CN)

Zhejiang Humanoid "NAVIAI" Claims Scaled Commercial Use Across Four Sectors · humanoid

Ningbo-based Zhejiang Humanoid says its fully self-developed "NAVIAI" series has already achieved scaled commercial use this year across four sectors: industrial manufacturing, commercial services, education/training, and data collection. This is a company self-report; specific installation numbers and revenue have not been independently disclosed.⚠️ Vendor-disclosed figuresSource: Ningbo Cyberspace Affairs source (WeChat, CN)

4. Industry Developments

South Korean Conglomerates Collectively Bet on Physical AI: Samsung, Hyundai Pledge ~₩312 Trillion in Southern Investments · industrial

South Korean giants Samsung, SK, and the Hyundai Motor Group unveiled industrial blueprints for the Yeongnam region on the same day, with combined investments of roughly ₩312 trillion (about $195 billion), covering AI data centers, robotics factories, and launch vehicles. Samsung alone will invest ₩60 trillion to build a physical AI manufacturing cluster in Yeongnam, plus ₩19 trillion for a robotics factory in Gumi; Hyundai announced roughly $27.3 billion in investment in the southeast to develop mobility and physical AI. Following Japan's "sovereign AI" ¥1 trillion push and China's embodied-AI surge, South Korea is now formally joining the physical AI race with state-level capital and manufacturing bases.Source: Reuters source, The Korea Times source, Chosunbiz source

Estun Plans Full Acquisition of Estun Cobot to Expand into Embodied AI · industrial

Chinese industrial robotics leader Estun Automation disclosed it is planning a full acquisition of its collaborative-robot subsidiary Estun Cobot, widely read by the market as a signal of consolidation into embodied AI, with the stock hitting the daily limit-up 3 times in 4 days. Industrial robot makers expanding into collaborative/embodied businesses is another instance in this cycle's trend of full-robot manufacturers filling gaps in their portfolios.Source: JRJ source

Unitree IPO Approval Ignites China's A-Share Robotics Sector, Over 40 Stocks Hit Limit-Up · humanoid

Following yesterday's approval of Unitree Robotics' IPO registration on the STAR Market (already reported), China's robotics industry index rose nearly 8% intraday to lead the market today, with Green Harmonic Drive and more than 40 other stocks hitting limit-up, and robotics ETFs saw large net inflows. Multiple brokerages view this as a "catalyst and industry validation" for the sector, though some analysts warn of divergence risk between "the tail end of thematic speculation" and "the early stage of earnings validation."Source: Cailianshe source, Zhongjin Online source

Boston Dynamics CEO: Without a National Strategy, U.S. Robotics Risks Repeating Semiconductors' Fate · humanoid

As China, Japan, and South Korea successively elevate physical AI to a national strategy, Boston Dynamics' CEO publicly called for the United States to adopt a national-level robotics strategy, warning that without one, U.S. leadership in robotics could suffer the same fate as semiconductors. This is an executive statement, but it stands in sharp contrast to South Korea's ₩312 trillion investment announced the same day, highlighting an emerging "industrial policy race" among nations.⚠️ Executive statementSource: finance.biggo.com source

CloudMinds Unveils Single-Arm Collaborative Embodied Robot · embodied

CloudMinds (Chinese robotics maker) launched its first single-arm collaborative embodied robot, positioned around fusing world models and VLA as complementary approaches (rather than a either/or choice), targeting hotel and commercial service scenarios. The product integrates "predicting the future + semantic understanding + action" into a single collaborative platform.Source: Ikanchai source

AgiBot Unveils "Event-Level Predictive" World Model WALL-WM · world-model

AgiBot Robotics (Chinese embodied AI startup) says it has released the "world's first event-level predictive" embodied world model, WALL-WM, arguing for moving beyond "learning pixel changes frame by frame" toward modeling task-semantic events instead. The "world's first" claim is a vendor self-description; independent comparisons and third-party reproduction are not yet available.⚠️ Vendor-disclosed claimSource: Dolphin Digital Intelligence Lab source (WeChat, CN)

Guangzhou and Shanghai Roll Out Embodied AI Policies and Real-World Pilot Programs · industrial

Local governments continue to ramp up support: Guangzhou issued "Measures for Promoting the High-Quality Development of the Embodied Intelligent Robotics Industry," covering data collection standards, intelligent computing, and open scenarios; Shanghai's Municipal Commission of Economy and Information Technology launched a 2026 real-world pilot program for humanoid robots and embodied AI, targeting the identification of more than 100 high-value application scenarios by year-end and deployment capacity at the scale of tens of thousands of units. These include forward-looking targets, but represent genuine institutional efforts driving the industry forward. Separately, China's Ministry of Human Resources and Social Security plans to add 12 new occupations, including "embodied AI robot application technician."Source: Guangzhou Industry and Information Technology Bureau source (WeChat, CN), KCSJ Daily source (WeChat, CN), Jiemian News source

Tesla's Optimus Production Line Transition; VP Says Mass Production Begins by Year-End · humanoid

Musk posted a photo of the Optimus factory, with reports indicating the Fremont production line is transitioning from Model S/X toward Optimus; a Tesla vice president said humanoid robot mass production would begin by year-end, targeting an annual output of one million units. The line transition is an observable fact, while the "one million units" figure is a target/plan; the mass-production timeline remains to be delivered.⚠️ Planned/target figuresSource: Dasheng Liaowang source (WeChat, CN), chinanews.com.cn source

Hardware · Supply Chain

· Dexterous hand motors: Multiple technical reviews note that dexterous hands are essentially "multi-axis micro motion-control systems" (with roughly 20 joints operating simultaneously); motors are upgrading from coreless designs to miniature servos with built-in drivers to compress wiring harness and controller size, making this a key bottleneck for scaling dexterous-hand production.

· Dense dexterous-hand funding: Industry tallies indicate the dexterous-hand sector saw 3 funding deals totaling more than ¥800 million in the first week of July; dexterous hands previously accounted for less than 5% of humanoid full-robot funding, but this imbalance began correcting starting in Q2, with hardware focus shifting from "brain/torso" toward "hands."⚠️ Third-party tally

· Inspire-Robots: Cited as the global benchmark for dexterous-hand mass production, with annual shipments exceeding 10,000 units, covering core structures for miniature lead screws, servo cylinders, and tactile dexterous hands, along with supporting force-control SDK and tactile-data-collection software.

· Harmonic drive reducers: A single humanoid robot requires more than twenty precision reducers, accounting for about 15% of total unit cost; as 2026 mass production ramps up, the industry expects harmonic drive reducer output to grow at a compound annual rate exceeding 85%, opening a window for Chinese-made substitution.⚠️ Brokerage forecast

FutureX · Physical AI Daily — Issue 46 (07/03)

Shawn — Thu, 02 Jul 2026 14:51:00 +0000

Today's Highlights

· Unitree Robotics' STAR Market IPO registration was approved (China Securities Regulatory Commission consent), planning to raise RMB 4.202 billion; at roughly 104 days from acceptance to registration effectiveness, this is the fastest pre-review on record for the STAR Market, putting China's "first listed humanoid robot company" on the horizon.

· World models are shifting from "contestant" to "infrastructure": D-Robotics' Uranus achieves frame-level closed-loop control, Zhiyuan Robotics (Chinese humanoid startup) released an open general-purpose training ground, and multiple world-model papers (Valdi, ABot-M0.5, etc.) landed the same day, extending the recent capital-attracting wave.

· Home embodied AI draws another major bet: Lexiang Technology (Chinese home robotics startup founded by former Dreame executives), completed a nearly RMB 500 million Pre-A round (led by Ant Group), bringing total funding to RMB 1 billion; Ant Group has now backed 12 humanoid/embodied AI companies in 18 months.

· Global bearing leader SKF formed a joint venture with Chinese harmonic reducer leader Leaderdrive, with SKF holding a 60% stake — moving directly into precision joint transmission for humanoid robots.

· The first Shanghai International Embodied Intelligence Expo opened; concurrently, the China Embodied Intelligence Industry Development Report (2026) stated China is among the world's fastest-growing embodied AI markets, with an annual compound growth rate of roughly 22%–23%.

I. Research Progress

"Domain arithmetic" adapts VLA to environmental drift with a single sample · vla

VLA models often fail wholesale when a camera pose changes or the robot changes (e.g., Panda→UR5e); traditional adaptation requires collecting multiple demonstrations per task, which is costly. This paper characterizes environmental drift as a "domain vector" that can be added or subtracted in weight space, achieving cross-domain adaptation with just a single sample — substantially reducing reliance on target-domain demonstration data. It topped the day's Hugging Face trending list with a score of 16.

Taewook Kang et al. · arXiv 2607.00666 source

Letting world models complete 3D point clouds for more stable long-horizon prediction · world-model

Large video dynamics models lack explicit 3D structure, causing geometric drift and error accumulation over long rollouts. This work injects 3D spatial structure into world models via point completion, outperforming 2D video baselines on both dynamics-learning accuracy and long-horizon consistency — aligning with the day's main theme of "world models moving toward usability."

Skand Peri et al. (Oregon State) · arXiv 2607.00148 source

A structured 4D latent prediction model for robot planning · world-model

Most video prediction models remain confined to 2D sequences, lacking 3D geometry and struggling with precise spatial reasoning. This work (including Yilun Du) builds a structured 4D (3D + time) latent prediction model, providing physically consistent future predictions for robot planning.

Zhiyi Li et al. · arXiv 2607.01166 source

ABot-M0.5: a world-action model unifying mobility and manipulation · world-model

Mobile manipulation remains difficult for current embodied learning approaches. ABot-M0.5 centers on "three-layer alignment" (temporal granularity, action space, and train-test consistency), using intermediate implicit actions to bridge video latents and embodied control — unifying mobility and manipulation within a single world-action model. Hugging Face trending score: 8.

Ronghan Chen et al. · arXiv 2607.00678 source · Analysis: Shuyuan AI source (WeChat, CN)

Valdi: a value-diffusion world model that jointly predicts value, reward, and dynamics · world-model

Diffusion models can capture uncertain future dynamics, but iterative inference is slow and ill-suited to low-latency latent-space planning. Valdi trains a latent diffusion dynamics model end-to-end within an online control loop, TD-MPC style, simultaneously predicting the value function, reward, and dynamics — balancing expressiveness with online usability.

Christopher Lindenberg, Kashyap Chitta · arXiv 2607.00917 source · Analysis: Shuyuan AI source (WeChat, CN)

Letting world models act as "referee": RoboWorld's neural simulation evaluates general-purpose policies · benchmark

Using video world models to evaluate general-purpose robot policies can bypass real-robot deployment, but the world model's own errors make rollouts unreliable, and inference is slow. RoboWorld targets policy evaluation that is both fast and reliable, echoing this week's discussion of "world models not as contestants, but as referees."

Byeongguk Jeon et al. (KAIST) · arXiv 2607.01060 source

Transferring human-hand tactile pretraining to dexterous manipulation · perception

Touch provides force feedback that vision cannot replace for contact-rich tasks, but tactile-equipped datasets are small in scale and narrow in contact coverage. This paper does human-centric, transferable tactile pretraining, alleviating the tactile data bottleneck in dexterous manipulation — resonating with the day's momentum around "tactile foundation models."

Chi Zhang et al. · arXiv 2607.01067 source

Fei-Fei Li, Jim Fan, and Danfei Xu: dexterous-hand data may have "gone down the wrong path" · manipulation

The three authors co-wrote a piece reflecting on the current approach to dexterous-hand learning, and released a 100-hour bimanual dexterous-hand teleoperation dataset covering over 200 everyday objects, 22 motion primitives (grasping, squeezing, insertion, wiping, folding, etc.), and more than 7,700 trajectories — arguing for reconstructing dexterous manipulation using data that better reflects human manipulation priors.

Fei-Fei Li, Jim Fan, Danfei Xu, et al. · Analysis: DeepTech source (WeChat, CN)

FurnitureVLA: a VLA for real-scale bimanual furniture assembly · vla

Existing robotic furniture assembly work mostly remains at toy scale or single-arm. FurnitureVLA is the first systematic study of a VLA for real-scale bimanual furniture assembly, pushing long-horizon two-handed manipulation closer to practical scenarios.

Chenyang Ma et al. · arXiv 2607.01212 source

Other papers today: FAR, enabling test-time recovery and continuous improvement via failure-aware retry (arXiv 2607.01111 source); ROSA, a foundation-model serving system for robot factories (arXiv 2607.01088 source); the paper for NVIDIA GEAR's self-improving skill-discovery system ASPIRE is now up (previously reported, arXiv 2607.00272 source); Wake up for Touch, mask-isolated tactile alignment in MLLMs (arXiv 2607.00302 source); From World Models to World Action Models, a tutorial survey of robot world models (arXiv 2607.00836 source); Path Planning in Physically Viable World Models (arXiv 2607.00673 source).

Open Source · Tools · Benchmarks

· Zhiyuan Robotics' General Training Ground: connects real-robot data, simulation evaluation, world models, and robot deployment into one open platform for community training and evaluation of embodied policies, aligning with the day's several "world model as evaluator" papers source (WeChat, CN)

· AgiBot's X-Tokenizer: a cross-modal embodied action tokenizer; the company claims a 13.5% improvement in multimodal alignment and an 8.25% improvement in long-horizon task performance, alongside a new QUANXTA Zero embodiment-free data collection solution source

II. Funding & Deals

Lexiang Technology (Zeroth) | Pre-A | Nearly RMB 500 million | RMB 1 billion total · embodied

Led by Ant Group. Lexiang was founded by former Dreame executives (Chinese cleaning-robot maker) just a year and a half ago; this round coincided with the launch of its home embodied-AI brand "Zeroth" (spanning collaborative, companion, humanoid, and other product forms). The company claims first-half revenue growth of roughly 600% and about 30,000 orders. Home scenarios are the new focus of the current wave of embodied AI capital.Source: Robotics Outlook source (WeChat, CN)

CarbonSix (South Korea) | Series A | $40 million (about KRW 60 billion) · industrial

Co-led by DSC Investment and LB Investment, with participation from KDB, IMM, and all existing seed-round shareholders. CarbonSix builds physical AI systems for factories (robotic intelligence software plus manipulators/end effectors), using a "data flywheel" approach that collects production-line data during deployment to feed back into the model; the company says it already has commercial contracts and revenue.Source: Chosunbiz source

Agility Robotics | SPAC listing | ~$2.5 billion valuation · humanoid

U.S. humanoid robot maker Agility plans to merge with Churchill Capital XI to go public under ticker AGLT, with the deal bringing in over $600 million in funding (including a PIPE led by Foxconn). Its bipedal robot Digit is already deployed at 9 customer sites including Schaeffler, GXO, and Toyota Canada, with over $300 million in multi-year orders on hand — positioning it as one of the few "pure-play humanoid" names on U.S. markets.Source: MarketBeat source

VAST | New round | Over RMB 1 billion · world-model

VAST, which started out in 3D generation, has raised over RMB 1 billion in a new round and is pivoting toward world models, aiming to use world models to support a low-barrier interactive content platform. Geely Capital invested in two embodied AI/world-model companies in a single day, with this round being one of them.Source: GeekPark source (WeChat, CN)

Kunlunxing Robotics | Three consecutive rounds | Tens of billions of RMB cumulative · embodied

Backed by Casstar (Zhongke Chuangxing), Hillhouse Ventures, Banyan Capital, Zhending Capital, and others. The company says it reached unicorn valuation within 90 days of registration, setting a new record for the fastest-growing embodied AI startup in China. The frequent appearance of such "unicorn-at-founding" narratives is itself a sign of overheating in the current sector.Source: Beijing Fund Industry Association source (WeChat, CN)

Luxonis | Series A | $14 million · hardware

Focused on the perception layer of physical AI (depth/spatial vision hardware and software). Compared with the recent cluster of "robot brain" funding rounds, this round directs capital toward the more upstream, less-covered sensing and perception stack.Source: The Robot Report source

III. Commercialization & Deployment

Pony.ai says its Guangzhou and Shenzhen Robotaxi operations have each reached per-vehicle profitability · autonomy ⚠️ Company claim

Pony.ai reported results claiming its Robotaxi operations in Guangzhou and Shenzhen have successively achieved per-vehicle profitability; the same day, the GAC Aion Hyper HT Robotaxi it backs began deliveries. Per-vehicle profitability is a key milestone for Robotaxi unit economics, but this is currently a company self-report, with methodology and depreciation assumptions not independently verified.Source: Autohome source

Blattner awards Built Robotics a $75 million contract · industrial

According to The Robot Report, energy engineering giant Blattner is deploying Built Robotics' physical AI robots at scale in U.S. solar farm construction under a contract worth roughly $75 million, handling high-risk tasks such as pile driving, surveying, and trenching, with operators supervising from a safe distance — pointing to real order flow driven by "clean-energy demand plus construction cost-and-safety improvements."Source: The Robot Report source

AgiBot's A3 humanoid lands in Europe, launches Robot-as-a-Service in the UK · humanoid

AgiBot debuted its A3 humanoid robot in Europe and launched a Robot-as-a-Service (RaaS) rental model in the UK. Following mass production at the ten-thousand-unit scale in China, leading manufacturers are beginning to treat overseas expansion and "rent-as-needed" as part of their scaling strategy.Source: Robotics & Automation News source

UBTECH humanoid robots deployed at a China border crossing for crowd control · humanoid

According to foreign media, UBTECH humanoid robots have been deployed at a busy international border crossing in China to assist with crowd flow/order management. Compared with home-companion use cases, this type of government/public-facing scenario is one of UBTECH's humanoid deployments that comes closer to real working conditions.Source: Interesting Engineering source

Novares partners with Innov8 to pilot humanoid robots in auto-parts manufacturing · humanoid ⚠️ Pilot-stage claim

Auto-parts maker Novares has partnered with Innov8 to pilot humanoid robots on an auto-parts production line. This is an early pilot, not a scaled deployment, and its real-world viability remains to be seen.Source: Ifeng Tech source

IV. Industry Developments

SKF and Leaderdrive form joint venture, entering precision joints for humanoid robots · hardware

Global bearing and transmission leader SKF and Chinese harmonic reducer leader Leaderdrive signed an agreement to establish a joint venture in China, with SKF holding a 60% stake, focused on high-precision transmission components for robot joints; operations are expected to begin by the end of 2026. The joint venture is anchored in China — the world's largest humanoid robot market — while leveraging SKF's global network to reach Europe, Japan, and the United States, marking a significant step by a foreign giant into humanoid joints, the "most expensive component."Source: PR Newswire source

D-Robotics' Uranus world model achieves frame-level closed-loop control · world-model

D-Robotics (a Horizon Robotics affiliate) released its Uranus world model, moving from "video generation" to "interactive simulation" with frame-level closed-loop capability. Unlike most manufacturers that position world models as an "embodied brain," Uranus is positioned as infrastructure for robot development (data generation, policy evaluation), echoing the day's industry-wide shift toward "world models as referees, not contestants."Source: Sohu source · QbitAI source (WeChat, CN)

Ant Group doubles down on humanoid robots: 12 deals in 18 months · adjacent

According to CNBC, Ant Group has invested in 12 humanoid/embodied AI companies since early 2025, with its lead investment in Lexiang Technology's nearly RMB 500 million Pre-A round the latest example. Capital from internet giants is accelerating into the application and full-robot layers of the embodied AI industry.Source: CNBC source

Beijing Humanoid Robot Innovation Center unveils "world's first full-scale, whole-body VLA technical framework" · humanoid ⚠️ Company claim

The Beijing Humanoid Robot Innovation Center says its "Wisdom Sail" (Huisi Kaiwu) platform has launched a full-scale, whole-body VLA framework, primarily addressing gait oscillation and motion instability issues that arise when mainstream VLA models are transferred to high-dimensional, whole-body humanoid control. The "world's first" claim is the company's own description; its real-world capability and production readiness remain to be verified.Source: Beijing Humanoid Robot Innovation Center source (WeChat, CN)

Musk shares group photo of Tesla Optimus mass-production team · humanoid ⚠️ Statement/claim

Musk posed with the Optimus production-line team at the Fremont factory (formerly producing the Model S/X, which ceased production in May), and acknowledged that humanoid mass production "will start out extremely slow, unlike car manufacturing." This reflects production-line conversion progress and commentary, with no mass-production delivery figures yet given.Source: Humanoid Robot Alliance source (WeChat, CN)

International Federation of Robotics (IFR) elects new president · adjacent

Jane Heffner of Teradyne was elected as the IFR's new president, completing a leadership transition at the industry governance level.Source: AI Insider source

Hardware · Supply Chain

· MEMS six-axis force sensor chip: Zhongke Midian says it has begun mass production of a MEMS six-axis force sensor chip, aiming to substantially cut the cost of six-axis force sensors — currently RMB 80,000–120,000 per unit, with each humanoid robot requiring 4–6 units at roughly RMB 20,000 each source (WeChat, CN)

· Joint-specific chip: China's first dedicated gallium-nitride magnetic encoder chip for humanoid robot joints has been released, claimed to cut power consumption by 22%, filling a domestic gap in high-end robot sensing source (WeChat, CN)

· Weihong Corporation: its micro servo motors have entered the North American dexterous-hand supply chain, becoming a "new favorite" among overseas manufacturers source

· Dexterous-hand valuation paradox: RealMan (Lingxin Qiaoshou), a Chinese dexterous-hand maker, has built a roughly RMB 43 billion valuation on the narrative of holding "over 80% global market share in high-DOF dexterous hands," while Yinshi Robotics, the actual shipment volume leader, has a comparatively modest valuation — capital is paying a premium for narrative, not for production volume source (WeChat, CN)

V. Weekly Watch

China rolls out a wave of humanoid robot standards and field-training policies · hardware

Regulators were highly active on humanoid robots this week: China's Ministry of Industry and Information Technology (MIIT) opened public comment on an industry standard for six-axis force sensors for humanoid robots, bringing this core force-control component into standardization; MIIT and the State-owned Assets Supervision and Administration Commission (SASAC) jointly launched a 2026 special initiative for humanoid robot and embodied AI field-scenario training, specifying that core products should see routine deployment across multiple typical scenarios by year-end; Hangzhou's Binjiang district and other locations also rolled out China's first humanoid robot quality-management standard (with Unitree and others participating in its drafting). The common thread: using standards and application scenarios to push the industry from "usable" to "good to use."Source: Home Service Robot Committee source (WeChat, CN) · Dajishi source (WeChat, CN)

Morgan Stanley: China's humanoid robot fleet could reach 1 billion units by 2050 · humanoid ⚠️ Single-firm long-range forecast

Morgan Stanley's China Humanoid Robot Industry report offers a long-range projection: by 2050, China's humanoid robot fleet could reach roughly 1 billion units, corresponding to a market of about $7.5 trillion; the bank also raised its forecast for China's 2026 humanoid shipment volume. This is a single investment bank's long-term scenario, using a different methodology from the "22%–23% compound annual growth" figure cited the same day in the China Embodied Intelligence Industry Development Report (2026) — best read as a directional trend rather than a firm number.Source: Robotics Knowledge Base source (WeChat, CN)

This week in supply chain: harmonic reducers see "foreign entry + Chinese duopoly" jockeying simultaneously · hardware

Harmonic reducers were the hottest supply-chain segment this week: on one side, a foreign giant is formally entering the market — SKF taking a 60% stake in a joint venture with Leaderdrive; on the other, China's duopoly is solidifying — Leaderdrive holds over 60% of the Chinese market and ranks second globally, while Laifu Harmonic just listed on the Hong Kong Stock Exchange as "the first harmonic reducer stock in Hong Kong" (shipping nearly 292,000 units in 2025). Industry data shows global harmonic reducer sales reaching roughly $850 million in 2026, with China's 2025 market at about RMB 1.6 billion and consumption of roughly 957,500 units — humanoid robot mass production is the main source of incremental demand.Source: Shenlian Industry Research source (WeChat, CN) · Industry Energy Circle source (WeChat, CN)

This week in supply chain: six-axis force sensors and robot chips — cost reduction versus the debate over "whether to build dedicated chips" · hardware

Six-axis force sensors remain the most prominent bottleneck in mass-production costs — each humanoid robot needs 4–6 units, at a total cost of roughly RMB 80,000–120,000 per unit — and this week saw parallel efforts to ease that via "MIIT's industry standard plus MEMS six-axis force chips lowering costs." On the chip side, automakers are collectively crossing over (Xpeng's Turing chip, MetaX's joint venture with UBTECH, ASR Microelectronics' 4nm partnership with D-Robotics) to jockey for position in joint control and compute; but Song Jiqiang, head of Intel China Research Institute, cautioned that now is not the best time to launch dedicated robot chips — VLA capabilities are still evolving, and general-purpose chips adapt better to algorithmic uncertainty — offering a note of caution amid the current "chip-building fever."Source: Bengbu Midian Technology source (WeChat, CN) · Gasgoo Embodied Intelligence source (WeChat, CN)

FutureX · Physical AI Daily — Issue 45 (07/02)

Shawn — Wed, 01 Jul 2026 14:48:58 +0000

Today's Highlights

· Japan establishes a national strategy for sovereign physical AI: up to ¥1 trillion (~$6.1 billion) in R&D funding over five years for a homegrown "physical AI" foundation model, targeting the deployment of 10 million robots across 18 industries including food service and healthcare by 2040, led by the SoftBank- and Sony-backed alliance Noetra.

· China State Taxation Administration data: sales revenue growth for embodied intelligence companies rose to 22.4% in the first five months of this year, while industrial enterprises' spending on embodied robots purchases increased 2.3x year-on-year, signaling the industry's move into scaled volume growth.

· The largest single angel round yet in the consumer embodied robotics space: DISCOVER Robotics (Qiuzhi Technology) — founded by Zhou Guyue, a former core member of DJI — raised over $100 million in its first funding round, with Legend Capital and Lenovo Capital among the investors.

· Nvidia's GEAR lab open-sources ASPIRE, a self-improving skill discovery system that uses an "execute–diagnose–repair–accumulate" closed loop to raise dual-arm handover success rates from 20% to 92%, while building up a reusable skill library.

· AVATR (Chinese automaker) obtains a China L3 autonomous driving test license, launching multi-scenario road tests on highways and expressways in Chongqing — another step forward for China's advanced autonomous driving governance framework.

I. Research Progress

AdaJEPA: Letting World Models Keep Evolving During Action · world-model

Latent-space world models are typically frozen after training, so once test-time distribution shifts degrade their predictions, planning fails along with them. AdaJEPA embeds test-time adaptation of the world model directly into an MPC control loop, requiring just one gradient step per replanning cycle to significantly boost planning success rate — a representative work bringing "world models that keep learning while in use" into closed-loop control. Authors include Yann LeCun and Mengye Ren.

Ying Wang et al. (NYU, et al.) · arXiv 2606.32026 source · Commentary: AIer Notes source (WeChat, CN)

MemLearner: Giving Video World Models "Contextual Memory" · world-model

Video world models lack memory, causing scene drift and inconsistency during long-horizon generation. MemLearner teaches the model to actively query contextual memory, directly reusing pretrained visual priors without training extra modules from scratch — ranking among today's top trending papers in the community (HF↑18).

Jiwen Yu et al. · arXiv 2606.31734 source

Position Paper: VLA's "Physical Reasoning" Cannot Be Verified · vla

A position paper aimed at cooling the VLA hype: the industry commonly interprets gains on manipulation benchmarks as evidence that "internet-scale semantic representations have transferred to physical execution generalization," but the authors argue this assumption has never been independently verified and cannot be tested under current evaluation protocols, calling for a shift toward a verifiable research paradigm. A pointed take that cuts at a real methodological weakness in current VLA evaluation.

Taozhao Chen et al. (University of Sydney) · arXiv 2606.30686 source

From Grasps to Dexterity: Large-Scale Grasp Data Can Also Train "Functional Dexterous Manipulation" · manipulation

Large-scale grasp data has previously stopped at grasp generation and pick-and-place. This paper pushes it further into functional dexterous manipulation that requires operating an object's internal mechanism — such as aiming a spray bottle to water plants or operating a glue gun. The method uses hierarchical imitation learning (a high-level module predicting hand sub-goals plus a low-level goal-conditioned controller), demonstrating that grasp data can serve as a scalable pretraining resource for contact-rich dexterous manipulation. From CMU's David Held group.

Ying Yuan et al. (CMU) · arXiv 2606.30749 source

World-Action Model Achieves First Sim-to-Real Transfer from Pure Synthetic Priors · world-model

Replacing costly real-robot demonstrations with scalable synthetic data has long been appealing, but world-action models had never before been shown to bridge the sim-to-real gap. This paper is the first to train a world-action model purely on synthetic priors and deploy it zero-shot to real-robot manipulation, opening a path for synthetic-data-driven manipulation learning.

Zixing Wang et al. · arXiv 2606.31101 source

SARL: Fine-Tuning Generalist Robot Policies via Semantic Reinforcement Learning · vla

Generalist robot policies make good priors for downstream RL, but directly RL fine-tuning them often destroys their generalization. SARL instead optimizes a semantic prompt space through online interaction, using the generalist policy as a steerable "skill prior" — significantly outperforming existing methods at improving deployed behavior. From UC Berkeley's Sergey Levine group.

Jagdeep Singh Bhatia et al. (UC Berkeley) · arXiv 2606.31958 source

Z-1: Efficient Reinforcement Learning for VLA from Its Own Failures · vla

Most VLAs rely on behavior cloning or supervised fine-tuning, with few chances to improve from their own failures. Z-1 uses systematic GRPO post-training to substantially improve flow-based VLA policies without needing additional private demonstrations, surpassing published SOTA models.

Lang Cao et al. · arXiv 2606.31846 source

Human-as-Humanoid: Turning Human Video Directly into Motion Supervision for Humanoids · vla

Collecting real-robot data at scale for high-DOF humanoids is extremely difficult. This framework jointly aligns robot embodiment, sensor setup, and action-label interfaces, converting first- or third-person human video near real-time into observation-action supervision usable by humanoids — making "learning humanoid control from human video" zero-shot feasible.

Xiaopeng Lin et al. · arXiv 2606.32009 source

Freeform Preference Learning: Learning Manipulation Policies from Freeform Human Preferences · manipulation

In long-horizon manipulation, sparse success labels are too weak a signal, while binary preferences collapse multiple quality dimensions into one blurry signal. FPL instead lets robots learn from freeform human preferences, improving 38 percentage points over sparse-reward and binary-preference baselines, and can steer policies toward different behaviors at test time without retraining. From Stanford's Chelsea Finn group.

Marcel Torne et al. (Stanford) · arXiv 2606.32027 source

Other papers today: DVG-WM (an efficient embodied world model with decoupled video generation, achieving better image quality on LIBERO and real robots with up to 3.97x speedup, arXiv 2606.32028 source); CoDex (zero-demonstration dexterous functional manipulation, using a VLM to infer semantic constraints then refining with RL, arXiv 2606.31909 source); World-Model Collapse as a Phase Transition (the implicit world model in long-horizon language agents collapses abruptly at a critical point, with world fidelity failing before action validity, arXiv 2606.31399 source); Delta-JEPA (latent difference-action decoding improves world models' sensitivity to actions, arXiv 2606.31232 source); RL control for humanoid roller-skating (from Marco Hutter's group, energy use drops about 50% versus normal gait when equipped with consumer roller skates, arXiv 2606.31807 source); a survey on Robustness of Robotic Manipulation (a unified conceptual and mechanistic framework for manipulation robustness, arXiv 2606.31494 source); Revisiting Parameter Redundancy in VLA (parameter evolution patterns and efficient pruning in VLM→VLA adaptation, arXiv 2606.31382 source); ChronoFlow-Policy (a diffusion visuomotor policy unifying past-present-future interaction flow, arXiv 2606.31493 source); One Video One World / OVOW (training-free reconstruction of a single monocular video into a simulatable, instance-level 4D mesh scene, arXiv 2606.31388 source); ForgeDrive (bidirectional cross-conditioning between vision and action, unifying simulation, planning, and visual odometry for autonomous driving, arXiv 2606.31226 source).

Open Source · Tools · Benchmarks

· NVIDIA ASPIRE (GEAR lab): a self-improving skill discovery system that automatically diagnoses and repairs code-as-policy programs from execution feedback, then accumulates verified fixes into a reusable skill library, raising dual-arm handover success from 20% to 92%; released alongside a batch of open Nvidia models and simulation libraries source

· XiaomiOneVL: an autonomous driving model officially released and fully open-sourced by Xiaomi, claimed to exceed explicit chain-of-thought approaches in accuracy with inference latency as low as 0.24 seconds source

· FluxVLA Engine: an embodied intelligence foundation stack open-sourced jointly by Alibaba Cloud and AgiBot (Chinese embodied robotics startup, formerly Zhiyuan) source

· StarVLA: an open-source framework for building VLAs in a modular, "Lego-block" fashion, covering training on multiple benchmarks through to real-robot deployment on a Franka arm source (WeChat, CN)

· WorldRoamBench: an open-world benchmark for the long-horizon stability of interactive world models, evaluating action following, memory, and interaction physics across four dimensions (arXiv 2606.31672 source)

II. Funding and Deals

DISCOVER Robotics (Qiuzhi Technology) | Angel Round | Over $100 Million · embodied

Backed by Legend Capital, Xinchan Investment, Lenovo Capital, and others. The company was founded by Zhou Guyue, a former core member of DJI, and incubated by Tsinghua University's Institute for Intelligent Industry; it develops fully self-developed, consumer-grade embodied robots. This round sets a record for the largest single angel round in China's consumer embodied robotics space, and reflects the day's broader capital enthusiasm for "building the brain upstream, deploying the body downstream." Source: Robot Outlook source (WeChat, CN)

Sinovation Motong | Series C | Nearly ¥200 Million · industrial

Led by Kingsway Investment, with Jiuxuan Capital as financial advisor; focused on smart equipment and production-line integration. Source: Sohu source

Ganzhi Jiyuan | Angel Round | Tens of Millions of RMB · hardware

Led by Sino-Rock (Songhe) Capital, with funds mainly going toward pilot-scale production line construction and product iteration. The company makes multimodal tactile electronic skin, entering the humanoid "tactile sensing" sub-sector that has recently attracted concentrated investor interest. Source: Robot Outlook source (WeChat, CN)

Variable-Stiffness Joint Startup Spun Out of Beihang University | Angel Round | Nearly ¥100 Million · hardware

Founded by a team from the Robotics Institute at Beihang University (Beijing University of Aeronautics and Astronautics), focused on intelligent variable-stiffness joints, a core upstream component for humanoid robots. Source: 36Kr source

Zhitianxia | Angel Round | Amount Undisclosed · world-model

A world-model startup describing itself as aiming to build "China's World Labs," riding the recent wave of investor interest in world models (previously reported). Source: LeiPhone source

Xijing Technology | ChiNext IPO Application Accepted · autonomy

A provider of autonomous driving solutions for commercial vehicles; its prospectus claims its related revenue scale ranks first in the industry. A wave of robotics and autonomous driving companies in China are rushing to file for listings on Hong Kong and mainland A-share exchanges. Source: Sina Finance source

III. Commercialization and Deployment

Amazon Deploys 5,000 Robots at New Superwarehouse in Poland · industrial

Amazon's newly built giant warehouse in Poland has put roughly 5,000 robots into operation, another large-scale expansion of its warehouse automation in Europe. Source: TVP World source

JD.com Scales Up Both Warehousing and Food-Service Deployments · industrial

At JD.com's "Asia No. 1" smart logistics park in Harbin, 84 "Geek+" warehouse robots are now running in routine operation; at the same time, JD.com is accelerating the scaled rollout of embodied robots in food-service settings, expanding from warehouse to storefront scenarios it owns. Source: NetEase/Sina Finance source

RideFlux and Hanjin Launch Commercial Autonomous Freight in South Korea · autonomy

South Korean autonomous driving company RideFlux has partnered with logistics leader Hanjin to launch commercial autonomous freight services — a step for South Korean trunk-line logistics autonomy moving from testing to commercial operation. Source: thelec.net source

Neolix L4 Autonomous Delivery Vehicles Obtain Test Permit in Malaysia · autonomy ⚠️ Test permit

Neolix (Chinese autonomous delivery vehicle maker) has obtained an autonomous delivery test permit in Malaysia, taking its L4 logistics vehicles into Southeast Asia; this is a pilot-access permit rather than a scaled deployment. Source: cheshi.com source

AgiBot's 15,000th Unit Delivered to Longqi Factory (Previously Reported) · humanoid

The 15,000th embodied robot (Lingjing G2) that rolled off AgiBot's line on June 28 was delivered the same day directly to a Longqi Technology factory, entering front-line smart manufacturing use. The production milestone and deployment progress were detailed in earlier coverage, noted here briefly. Source: AGI Research Society source (WeChat, CN)

IV. Industry Developments

Japan Bets on "Physical AI" as a National Priority: ¥1 Trillion Over Five Years, 10 Million Robots by 2040 · world-model

Japan has formally established a national strategy for sovereign physical AI: the Ministry of Economy, Trade and Industry and its innovation agency have commissioned the Noetra alliance — led by SoftBank and Sony — along with the national research institute AIST, to develop a homegrown "physical AI" foundation model over fiscal years 2026–2030, with up to ¥1 trillion (~$6.1 billion) in funding over five years. The strategy's goal is to deploy 10 million robots across 18 industries including food service, food manufacturing, and healthcare by 2040, easing the effects of an aging population and labor shortages; Minister of Economy, Trade and Industry Ryosei Akazawa said the ministry would "go all-in on driving social implementation." This makes Japan the latest major economy, following China and South Korea, to elevate embodied/physical AI to the level of a national foundation-model priority. Source: Khaleej Times source

China's MIIT and SASAC Launch 2026 Special Action on "Real-World Field Training" for Humanoid Robots · humanoid

China's Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission jointly issued a notice launching a 2026 special action on real-world field training for humanoid robots and embodied intelligence, specifying that by the end of 2026, core humanoid robot products should move into routine, real-world operational use across multiple typical scenarios. Combined with the industry standard YD/T 6770-2026 ("Benchmarking Methods for Embodied Intelligence"), which took effect June 1, China's embodied intelligence sector is entering a policy phase of "standards to follow, field training to execute." Source: Daji Shi source (WeChat, CN)

AVATR Obtains China L3 Autonomous Driving Test License · autonomy

AVATR (Chinese automaker) has obtained a China L3-level autonomous driving test license, launching multi-scenario road tests on highways and expressways in Chongqing. Following the UN's release of a global technical regulation on automated driving and China's push to submit mandatory national standards for L3/L4, China's L3 autonomous driving governance framework is progressively coming together, marking another step toward mass-production deployment. Source: The Beijing News (via DoNews) source

South Korea Releases Physical AI R&D Blueprint, Boosts National Fund · world-model

Following South Korea's elevation of physical AI to a national strategy (previously reported), the Ministry of Science and ICT has released an R&D blueprint proposing to overcome physical AI's data bottleneck through domestically developed world models. The national growth fund is simultaneously increasing related investment, and the government is accelerating the establishment of a safety certification system for humanoid robots. Source: The Korea Herald source

MEIL and Analog Plan $500 Million Push to Bring Physical Intelligence to India · adjacent ⚠️ Plan stage

Indian infrastructure conglomerate MEIL and UAE-backed Analog plan to invest approximately $500 million to deploy next-generation physical intelligence and AI infrastructure in India.Source: Business Standard source

World Model Momentum Continues, Academia Moves to "Set the Record Straight" · world-model

World models remain the day's dominant theme (previously reported): the 2026 Beijing Academy of Artificial Intelligence Conference sought to clarify the muddled definition of "world models," Fei-Fei Li wrote that world models are the gateway to physical intelligence, and both capital and academia continue to build momentum around the idea of a "robot brain." Source: Shake Network Technology News source (WeChat, CN)

Nvidia Expands Robotics Teams Across Three Locations in China · adjacent

Nvidia is expanding its robotics teams in three locations in China, positioning itself as providing "not a complete robot, but an Android-like operating system for robots" — i.e., supplying the platform and foundation software rather than the hardware itself, reinforcing its foothold in the embodied AI ecosystem's lower layers. Source: Sina Finance source

Hardware · Supply Chain

· Dexterous hands: Digitimes reports that price competition in China has cut the cost of humanoid robot hands by roughly half, though precision components such as six-axis force sensors remain difficult to bring down in cost at the same pace; GGII data projects China's dexterous hand sales will reach 70,000 units in 2026, about 3.6 times the 2025 figure source

· Joyson Electronics (Chinese Tier-1 auto parts supplier): to launch robotic dexterous hands and semi-solid-state battery solutions, as Tier-1 auto parts makers move in bulk into the embodied robotics supply chain source

· LiDAR: rising physical AI demand drove Ouster and Aeva shares up about 14% and 11% respectively in a single day source

FutureX · Physical AI Daily — Issue 44 (07/01)

Shawn — Tue, 30 Jun 2026 14:58:28 +0000

Today's Highlights

· Uber and Waymo end nearly three years of Robotaxi partnership in Phoenix: Waymo's fleet shifts to DoorDash delivery and Via transit, while Uber pivots to Lucid and Nuro with plans to deploy 20,000 vehicles over six years; both companies continue their partnership in Austin and Atlanta.

· UBTECH Robotics (Chinese humanoid robot company) launches full-size ultra-bionic humanoid robot U1 under consumer brand "UWORLD," priced at RMB 119,800–990,000 with 88 degrees of freedom, targeting home emotional companionship; pre-sale orders exceed 11,000 units (⚠️ pre-sale figures).

· World model startup funding wave continues: Forbes reports the sector has raised billions of dollars this year; today Tsinghua-affiliated Liqi Intelligence closed a seed round of hundreds of millions of RMB, while Dexmal completed a RMB 1 billion Series B targeting an IPO (valuation inherited from Zibianliang and Zhipingfang at RMB 20 billion).

· Apptronik opens an approximately 90,000 sq ft "Robot Park" training facility in Austin and unveils Apollo 2 (bipedal + wheeled), partnering with Google DeepMind to collect real-world data for training humanoid AI.

· Laifual Drive (Chinese harmonic drive manufacturer) lists on the Hong Kong Stock Exchange as the first "harmonic reducer stock" on the Hong Kong market, raising approximately HKD 1.07 billion; one of only two Chinese manufacturers to have achieved mass production of harmonic reducers for humanoid robots (⚠️ three consecutive years of losses).

I. Research Papers

DreamForge-World 0.1: A Low-Compute, Real-Time Interactive World Model Preview · world-model

World models are typically constrained by offline generation and high compute demands. This paper delivers a controllable real-time interactive world simulation in a low-compute preview: it adds a residual action pathway on top of the LongLive autoregressive video stack (derived from Wan2.1-T2V-1.3B), drawing on Matrix-Game concepts to achieve "feed action, generate future frames simultaneously." It ranks as today's top community paper (HF 7↑), representing the engineering direction of world models shifting from offline long-video generation toward interactive, low-latency systems.

Daniyel Ayupov et al. · arXiv 2606.30292 source

Orca: Unifying Multimodal World Signals into a "Universal World Foundation Model" · world-model

Current world models tend to be built separately per task or modality. Orca attempts to learn a unified world latent space, acquiring shared representations from multimodal world signals and exposing them via multimodal readout interfaces — pointing toward a world foundation model paradigm of "one base, many downstream uses."

Yihao Wang et al. · arXiv 2606.30534 source

Heterogeneous Tactile Transformer: One Representation to Bridge Heterogeneous Tactile Sensors · perception

The biggest pain point in tactile data is that sensors are incompatible — a model trained on one sensor fails when transferred to another, making large-scale data aggregation difficult. HTT learns shared representations across heterogeneous tactile sensors, allowing contact data from different tactile hardware to be pooled for training. This is the algorithmic counterpart to today's wave of tactile data infrastructure investment (Qianjue, South China University of Technology, Weitai).

Jianxin Bi et al. · arXiv 2606.29948 source

Sim-to-Real Physical Modeling for Professional-Level Robot Table Tennis · manipulation

High-speed spinning table tennis trajectories are counterintuitive; a robot must track and return the ball precisely within fractions of a second, while real-environment training is both expensive and hazardous. The Sony AI team builds a high-fidelity physical model, trains reinforcement learning policies in simulation, then transfers to physical robots, targeting professional-level high-speed gameplay — a hard sim-to-real benchmark for dynamic, contact-constrained control tasks.

Christian Conti et al. (Sony AI) · arXiv 2606.28805 source

Human2Any: Constraint-Aware Compositional Transfer from Human to Robot · manipulation

Human video is a scalable source of manipulation supervision, but morphology gaps, scene variation, and robot feasibility constraints make direct transfer difficult. Human2Any uses constraint-aware compositional planning to translate human demonstrations into actions executable by different robots, mitigating the embodiment gap; co-authors include NVIDIA researchers.

Shuo Cheng et al. (incl. NVIDIA) · arXiv 2606.28813 source

AnyBody: Whole-Body Humanoid Control Driven by Any Keypoint Subset · locomotion

AnyBody proposes a unified whole-body humanoid controller that can be commanded at deployment time by any subset of body keypoints — specifying only hand, foot, or torso targets is sufficient to generate coordinated whole-body motion, improving flexibility for teleoperation and motion retargeting.

Shuning Li et al. · arXiv 2606.29209 source

Trust Your Instincts: VLA with Test-Time RL Using Its Own Confidence · vla

Applying reinforcement learning to vision-language-action models typically requires external environment feedback and predefined success signals. This paper lets the model use its own confidence as an intrinsic signal for test-time RL, enabling online self-improvement without external success judgments — lowering the deployment barrier for VLA reinforcement learning.

Siyao Chen et al. · arXiv 2606.29892 source

You Only Touch Once: 6-DoF Object Pose Estimation from a Single Contact · perception

Visual pose estimation often fails with occlusion, reflective surfaces, or transparent objects. YOTO recovers the full 6-DoF pose of an object from just a single simultaneous contact pair, without requiring contact history, providing a purely tactile pose source for contact-rich manipulation.

Pengfei Ye et al. · arXiv 2606.28899 source

Other papers today: Flow Matching in Feature Space for Stochastic World Modeling (authors include INRIA, Meta FAIR, and others); SA-VLA (state-aware action tokenizer improving discrete action reconstruction accuracy); OWMDrive (4D occupancy world model for causally-aware end-to-end driving); WARP (whole-body retargeting from offline human demonstrations, extending mobile manipulation); TacGen (vision-to-tactile alignment and generation to address tactile data scarcity); J-LAW (coupled latent factor graph for joint localization and actionable world modeling); RoboGaze (structured vision-language analysis for evaluating robot world model-generated video); X-Mind (predictive world CoT for Xpeng's end-to-end driving, already reported on the industry side).

Open Source · Tools · Benchmarks

· Ruka-v2: NYU releases a fully open-source tendon-driven dexterous hand covering core degrees of freedom including wrist and finger abduction, aiming to replace "commercial hands costing tens of thousands of dollars" with a low-cost reproducible alternative, lowering the barrier for dexterous manipulation research.

· Qwen-AgentWorld: Alibaba open-sources an agent-oriented "language world model" that lets AI mentally simulate the consequences of an action in language before deciding on the next step.

II. Funding & Deals

Laifual Drive (03952.HK) | Hong Kong IPO | Raised ~HKD 1.07 Billion · hardware

Zhejiang-based Laifual Drive (Chinese harmonic reducer manufacturer) listed on the Hong Kong Stock Exchange on June 30, becoming the first "harmonic reducer stock" on the Hong Kong market, with an IPO price of HKD 85.5 and net proceeds of approximately HKD 1.073 billion, exclusively sponsored by CMB International; funds will be used for capacity expansion and R&D. By 2025 shipment volume, it ranks second among harmonic reducer makers for robots in China with approximately 21.4% market share, and is one of only two Chinese manufacturers to have achieved mass production of harmonic reducers for humanoid robots. 2025 revenue was RMB 261 million (up 142% year-on-year) but with a net loss of approximately RMB 171 million — another example of robot component makers pursuing a "sacrifice margins for volume, rush to capital markets" strategy.Source: Jiazi Guangnian (Chinese tech media) source (WeChat, CN)

Dexmal | Series B | RMB 1 Billion | Valuation Exceeds RMB 10 Billion · embodied

Shenzhen-based embodied AI company Dexmal (Chinese embodied AI startup) has closed a RMB 1 billion Series B, joining the RMB 10 billion valuation club and eyeing an IPO; Shenzhen Capital Group invested in two consecutive rounds, and Lens Technology (founder Zhou Qunfei) upgraded from customer to shareholder. The company's H1 2026 revenue is projected to approach RMB 100 million, with a full-year target of RMB 250–300 million, focusing on foundational world models, physics engines, and humanoid robot deployment — one of the few embodied AI unicorns to put actual revenue on the table.Source: Pencil News (Chinese startup media) source (WeChat, CN)

Liqi Intelligence (Tsinghua-affiliated) | Seed Round | Hundreds of Millions of RMB · world-model

Tsinghua University-affiliated world model startup Liqi Intelligence (Chinese AI startup) has closed a seed round of hundreds of millions of RMB, with Shunwei Capital, Sequoia China, Hillhouse, and Xinglian all participating. The team deliberately downplays the "world model" label, emphasizing coupling data, models, hardware, and infrastructure into systems that actually work in real scenarios — reflecting investors' preference for the "physics + data dual-flywheel" approach.Source: 36Kr source (WeChat, CN)

Yisheng Technology (founded by HKU professor) | Angel Round | Hundreds of Millions of RMB · adjacent

Yisheng Technology (Chinese AI startup), founded by a University of Hong Kong professor, has raised hundreds of millions of RMB in an angel round, focused on "building a memory system for robots" — addressing the long-neglected memory and long-horizon consistency challenges in embodied intelligence.Source: 36Kr Hardware (Chinese tech media) source (WeChat, CN)

NeoWa Robotics | Angel Round | RMB 50 Million · embodied

NeoWa Robotics (Chinese embodied AI startup) has closed a RMB 50 million angel round led by Lanhu Capital with Butong Capital and Gongqingcheng Puyi following; just two months ago it closed a seed round led by Plug and Play China; the founder is the former head of Baidu's autonomous driving and robotics lab, focusing on "embodied intelligence / universal traversal models." Seven other Shanghai companies also closed new rounds on the same day, reflecting continued dense early-stage deal flow in the embodied AI sector.Source: Tech Capital Circle (Chinese media) source (WeChat, CN)

III. Commercial Deployment

Amazon Unveils Next-Generation Warehouse Robot Proteus with Autonomous Package Handoffs · industrial

Amazon has revealed its next-generation autonomous mobile robot Proteus, featuring voice command support and enabling fully autonomous package handoffs from one robot to the next stage in the warehouse. As the world's largest-scale warehouse robotics operator, Amazon's iteration signals a shift in warehouse automation from "point-solution sorting" toward "continuous unmanned material flow."Source: MSN/Fox 59 source

HAI Robotics "Flash Climb" Series Surpasses 10,000 Units in Global Partnerships · industrial

Warehouse robotics company HAI Robotics (Chinese warehouse robot maker) reports its "Flash Climb" series has exceeded 10,000 units in global partnerships, including a collaboration with Arvato to serve a European beauty retailer's omnichannel smart warehouse. Scale has become the moat in box-handling warehouse robotics, though "partnership scale" is a cumulative figure rather than current installed base (⚠️ vendor figures).Source: Ikanchai source

Zhiyuan Robotics' 15,000th G2 Unit Delivered Straight to Factory (Previously Reported) · humanoid

Zhiyuan Robotics' (Chinese humanoid robot company) 15,000th embodied robot Elf G2 rolled off the production line and was delivered the same day to Longqi Technology's factory for production-line quality inspection — less than three months after surpassing 10,000 units in March. This story was already highlighted on June 28; today's follow-up adds supply chain signals including the inauguration of a Tsinghua–Zhiyuan joint research center and the adoption of GaN technology for joint drives at scale.Source: Shangguan News source

IV. Industry Developments

UBTECH Robotics Launches Full-Size Ultra-Bionic Humanoid U1, Priced at RMB 119,800–990,000 · humanoid ⚠️ Pre-sale figures

UBTECH Robotics unveiled its consumer brand "UWORLD" and debut full-size ultra-bionic humanoid robot U1 at its 2026 global launch event, displaying over 50 appearance variants, heights of 1.6–1.85 m, and 88 degrees of freedom across the whole body, targeting home emotional companionship. Pricing ranges from RMB 119,800 for U1 Lite and RMB 169,800 for U1 Pro, to RMB 990,000 for U1 Ultra (male version) and RMB 880,000 (female version); pre-sales launched on JD.com on June 2, with cumulative pre-orders exceeding 11,000 units as of the launch date (pre-sale orders, not deliveries). Founder Zhou Jian stated that robots will replace smartphones as the core AI interaction terminal, with half of the company's energy now directed at home scenarios. This represents the most aggressive attempt yet by China's humanoid sector to expand from "industrial/research" to "consumer companionship," though mass-production delivery and real-world retention remain to be proven.Source: Shanghai Securities News source (WeChat, CN)

Uber and Waymo End Phoenix Robotaxi Partnership, Each Pursues New Direction · autonomy

After nearly three years of operation, Uber and Waymo have ended their Phoenix ride-hailing partnership (the food delivery collaboration had already ended in May 2025). Waymo is reclaiming these vehicles into its own fleet to fulfill autonomous delivery commitments with DoorDash and transit agreements with Via; Uber is pivoting to Lucid + Nuro, planning to deploy more than 20,000 Lucid vehicles equipped with Nuro Driver over six years. Both companies continue their partnership in Austin and Atlanta, and Uber says a new Phoenix partner announcement is coming soon. This marks a shift in the Robotaxi model from "aggregation platforms + third-party fleets" toward vertically integrated control of vehicle capacity.Source: Reuters et al. (×30) source

Apptronik Opens Austin "Robot Park" and Unveils Apollo 2 · humanoid

Apptronik opened its approximately 90,000 sq ft flagship data collection and training facility "Robot Park" on June 30, where fleets of humanoids work on real logistics, manufacturing, and retail tasks to generate training data; it also unveiled humanoid platform Apollo 2 (in bipedal and wheeled variants), featuring 90%+ energy-efficient proprietary actuators, hot-swappable batteries for continuous operation, and expressive interaction. Data will be used to train future humanoid AI models in collaboration with Google DeepMind. The company, which raised $520 million at a $5 billion valuation in February this year, is now materializing its "robot data flywheel" and planning to replicate the Robot Park network across multiple locations.Source: The Star source

YOUIBOT Launches Industrial Embodied Foundation Model FabriX and Native Humanoid "Xifeng" · industrial ⚠️ Vendor figures

YOUIBOT (Chinese industrial robot company) globally launched a new series of embodied intelligence products, including what it calls the "world's first scalable" industrial embodied intelligence foundation model "FabriX (Zhihe)" and the industrial-native humanoid robot "Xifeng," claiming it will empower 10,000 industrial sites within three years. At a time when general-purpose humanoids are crowding into consumer narratives, industrial settings are seen as the most commercially viable near-term deployment ground; however, the "10,000 sites" figure represents cumulative empowerment including integration projects and should be distinguished from actual humanoid unit installations.Source: IItime (Chinese industry media) source (WeChat, CN)

LG Electronics Places Robotics Business Center Under Direct CEO Oversight, Bets on Physical AI · adjacent

LG Electronics has established a Robotics Business Center reporting directly to the CEO, consolidating dispersed operations and accelerating commercialization of physical AI and its home robot "Cloid." Following Samsung and Hyundai, LG is elevating robotics to group-strategy level, consistent with South Korea's broader move to designate physical AI as a national strategic priority.Source: The Korea Times et al. (×4) source

Nvidia Accelerates Hiring for Its China Robotics Team · world-model

Nvidia has posted 12+ new positions for its robotics team in China, interpreted as a talent signal of continued investment in its physical AI and embodied intelligence platforms (GR00T, Cosmos, and others).Source: South China Morning Post source

Hardware · Supply Chain

· STAR ROBOTICS (Chinese robotics startup) launches a 21-DoF fully direct-drive dexterous hand, with motors emerging as a key competitive differentiator — dexterous hands are evolving from mechanical components to highly integrated systems combining micro-motors, precision drives, encoders, drivers, and control algorithms.

· Sonair launches what it claims is the world's first safety-certified 3D ultrasonic sensor (ADAR) for human-robot collaboration near-field collision avoidance, adding a low-cost safety sensing pathway alongside vision and LiDAR (×5 multi-source).

· Dexterous hand valuations are approaching those of Unitree: Lingxin Qiaoshou's (Chinese dexterous hand startup) latest target valuation is approximately $6 billion (around RMB 41 billion), claiming over 80% global market share in high-DoF dexterous hands (⚠️ vendor figures); upstream components including miniature ball screws and harmonic reducers are also heating up simultaneously.

FutureX · Physical AI Daily — Issue 43 (06/30)

Shawn — Mon, 29 Jun 2026 14:47:45 +0000

Today's Highlights

· Shenzhen-based AI²Robotics (Chinese humanoid startup) and X Square Robot (Chinese embodied AI startup) both crossed the 20 billion RMB valuation mark on the same day: AI²Robotics raised nearly 5 billion RMB in a new round, while X Square Robot completed four rounds in two months with a reported valuation exceeding 20 billion RMB — bringing China's embodied AI "billion-yuan club" to approximately 25 companies.

· Volkswagen is reportedly moving to terminate its autonomous driving partnership with Bosch — after roughly €1.5 billion invested over four years, an internal assessment deemed the technology "not competitive," and the company is pivoting to external suppliers with a target of signing a new contract before September.

· South Korea has elevated physical AI to a national strategy: the government plans to invest 20 trillion KRW in manufacturing AI, commercialize industry-specific humanoid robots across 10 sectors by 2028, and has joined 11 companies to establish a Physical AI Industry Association.

· Tesla is rolling out FSD 14 Lite to owners of older HW3 vehicles, though those vehicles remain ineligible to join Tesla's robotaxi fleet.

· Multiple world model announcements landed on the same day: Xpeng unveiled predictive world model framework X-Mind at CVPR 2026 (in service of VLA 2.0), while Wujie Power (Chinese robotics startup) released its "latent-space world model MWA," claiming top benchmark rankings (⚠️ vendor claim).

I. Research Papers

PhysisForcing: A Physics-Constrained Reinforcement Approach for Robotic Manipulation World Simulators · world-model

Addresses a core limitation of video-generative world models for embodied manipulation: both general-purpose video generators and robot-data fine-tuned models produce physically implausible outputs — discontinuous trajectories, interpenetration between robots and objects — undermining their reliability as world simulators. This work injects physical consistency into the generation process via reinforcement, improving the physical plausibility of manipulation rollouts. HF↑30 on the day of release.

Peiwen Zhang et al. · arXiv 2606.28128 source

Using "Translation" as a Bridging Action: Transferring Human Manipulation Skills to Bimanual Robots · manipulation

Human motion data is cheap, abundant, and diverse — the most promising resource for scaling robot learning — but the morphological gap between human hands and parallel grippers makes direct transfer difficult. This work uses a class of "translation" bridging actions to map human demonstrations onto bimanual robots with parallel grippers, demonstrating that new manipulation skills can be learned from human motion data. HF↑27 on the day of release.

Sijin Chen et al. · arXiv 2606.28133 source

Injecting a Single Grounded 3D Anchor Point into the Action Head Unlocks Spatial and Task Generalization in VLAs · vla

VLA models achieve flexible manipulation through large-scale vision-language pretraining, but remain brittle along two axes at test time: object positions outside the training distribution (spatial generalization) and familiar scenes paired with new instructions (task generalization). This work finds that simply feeding a single 3D-grounded spatial point directly into the action head simultaneously improves both types of generalization — a minimal intervention with targeted effect.

Shiang-Feng Tsai et al. · arXiv 2606.27663 source

S²-VLA: State-Space-Guided VLA for Long-Horizon Manipulation · vla

Existing VLAs commonly suffer from perceptual drift and intent loss in long-horizon tasks due to accumulating errors. S²-VLA uses state-space-guided adaptive attention to suppress error accumulation; the team reports that with 2B parameters it outperforms 7B-scale baselines on multiple long-horizon manipulation benchmarks.

Zhipeng Xie et al. (Shanghai Jiao Tong University, East China Normal University) · arXiv 2606.27872 source · Commentary: Embodied Intelligence Notes source (WeChat, CN)

Drop and Recover: How Redundant Are VLA Models? · benchmark

VLAs directly inherit the oversized language backbones of pretrained VLMs — capacity far exceeding what is needed to process short robot instructions. This study uses a "drop-and-recover" probe to systematically measure how much model capacity closed-loop control actually requires, providing quantitative grounding for VLA pruning and compression with direct relevance to on-device deployment.

Guoheng Sun et al. · arXiv 2606.27755 source

SceneBot: Contact-Prompted Whole-Body Humanoid Tracking and Scene Interaction · locomotion

Current humanoid RL policies handle free-space motion well but struggle with contact-rich tasks — pure kinematic tracking cannot resolve the physical ambiguities that arise when interacting with objects or uneven terrain. SceneBot uses contact prompts to unify free-space walking, terrain traversal, and whole-body manipulation within a single motion-tracking framework.

Sirui Chen et al. (Guanya Shi group) · arXiv 2606.27581 source

Perceptual 3D Simulation: Predicting Scene Evolution via Physical World Modeling · world-model

"Given an image, predict how a scene evolves under a 3D transformation" is a shared goal across vision, graphics, and robotics — yet real systems, unlike ideal simulators, cannot fully access 3D geometry and dynamics, and must rely on partial, incomplete perceptual inputs. This work models physical evolution under perceptual constraints, advancing toward actionable, reality-level simulation.

Wanhee Lee et al. (Stanford, Daniel Yamins group) · arXiv 2606.27575 source

Other papers today: SpikeVLA (low-power real-time VLA using spiking neural networks, arXiv 2606.27807 source); Support-Constrained RL (improving real-world policies without real-robot experience, arXiv 2606.27475 source); Unleashing Infinite Motion (amplifying expressive quadruped locomotion with generative video priors, arXiv 2606.28237 source); DexCompose (reusing dexterous manipulation policies for single-hand multi-task execution, arXiv 2606.28323 source); Booster Lab (data-centric pipeline for deployable humanoid locomotion, arXiv 2606.27813 source); DIM-WAM (world-action modeling with diverse historical event memory, arXiv 2606.27677 source); ReWorld (learning better representations for world-action models in autonomous driving, arXiv 2606.27504 source).

Open Source · Tools · Benchmarks

· MemoBench: A world-modeling benchmark focused on dynamically changing environments — specifically testing memory consistency when targets move out of view or are occluded while the scene continues to evolve, filling a gap left by most benchmarks that only test targets within the field of view (arXiv 2606.27537 source).

· SimFoundry: A modular, automated scene generation system capable of zero-shot real-to-sim reconstruction from a single video, for use in policy learning and evaluation (arXiv 2606.28276 source, HF↑6).

II. Funding & Deals

X Square Robot (Chinese embodied AI startup) ｜ B+/B++/C Rounds ｜ Post-money valuation exceeding 20 billion RMB (~$2.8B) · embodied

Three consecutive rounds — B+, B++, and C — all closed within just over two months, making it the first company in the Greater Bay Area to reach a 20 billion RMB valuation. The investor lineup is an unusually complete roster of Chinese internet majors: Meituan, Alibaba, ByteDance, and Xiaomi each led different rounds, with Sequoia China, IDG, Source Code Capital, the National AI Industry Fund, and China Mobile also participating. The company focuses on embodied foundation models and integrated robots, with deployments already live in real-world settings including 58 Daojia (Chinese home services platform) and a component production line at a German luxury automaker. Source: Pencil News source (WeChat, CN)

AI²Robotics (Chinese humanoid startup) ｜ New Round ｜ ~5 billion RMB (~$700M) ｜ Valuation exceeding 20 billion RMB · embodied

Announced its 20 billion RMB valuation on the same day as X Square Robot. This round was jointly led by the National SME System Fund, the Guangdong Provincial AI Fund, Shenzhen Capital Group (深创投), Nanshan Strategic Emerging Industry Investment, and strategic investors including Moutai Group (Chinese spirits conglomerate), China Merchants Capital, and Sino Biopharmaceutical — with the valuation doubling from the ~10 billion RMB Series B in February. The company positions itself around general-purpose productivity robots led by AlphaBot2 and a proprietary brain-inspired VLA, with its own production line now at hundreds of units per month and a target of reaching tens of thousands this year. Source: RecodeX source (WeChat, CN)

SpeedBot (Chinese industrial robot vision startup) ｜ Series B++ ｜ 100M+ RMB range · industrial

This round was co-invested by the National SME Hunan Sub-Fund managed by Dachen Caizhì (Chinese PE firm) and Changsha Urban Development. The company has previously completed seven rounds, including a Series B of 300 million RMB. SpeedBot uses an industrial world model to build physical cognition for manufacturing scenarios, supporting embodied AI deployment on production lines. Source: Changsha Enterprise Listing Service Center source (WeChat, CN)

Delta Intelligence ｜ Angel+ Round ｜ Undisclosed · world-model

A humanoid robot foundation model company, this round was backed by Zhuhai Science and Technology Industry Group, Haiwan Capital (Chinese VC), Meihua Ventures (Chinese VC), semiconductor investor Jiaoyuan Assets, and a strategic automotive industry participant; existing investor Huaying Capital (Chinese VC) followed on. Source: Tech Capital Circle source (WeChat, CN)

Proception ｜ New Round ｜ $11 million · embodied

A Y Combinator-backed robot dexterous hand startup whose founder Jay Li previously served as technical lead for Tesla's Optimus program. The company announced this funding round simultaneously with the settlement of a trade secret lawsuit brought by Tesla (Tesla withdrew the suit this month). Source: TechCrunch source

Archon Robotics ｜ Seed Round ｜ Hundreds of millions of RMB · embodied

Positioned as a general-purpose whole-body embodied intelligence developer, the company has completed a seed round in the hundreds of millions of RMB. Source: Gasgoo source

Qingyu Technology ｜ Pre-Series A ｜ Undisclosed · embodied

A portfolio company of Shunwei Capital (Chinese VC fund backed by Xiaomi's Lei Jun), focused on tendon-driven soft robots — a direction widely regarded as one of the most difficult embodied robotics approaches to bring to production. The company claims its first-generation bionic flexible product is among the few tendon-driven systems globally to reach mass production. Source: Shunwei Capital source (WeChat, CN)

III. Commercial Deployment

Shenzhen Further Deregulates Robotaxi Commercialization · autonomy

Reports indicate Shenzhen will advance the legalization of commercial robotaxi operations starting July 1. As of 2026, 26 cities across China have been approved to operate driverless robotaxi services with commercial fares, with several jurisdictions progressively removing the requirement for a safety operator in the vehicle. Industry discussion has shifted from "can the technology operate" to questions of operations and per-vehicle unit economics. Source: AI Dimension Bureau source (WeChat, CN)

Aurora's Autonomous Trucking Safety Case Endorsed by Third-Party Audit · autonomy

An independent third-party audit has endorsed Aurora's autonomous trucking safety case — a materially more credible safety signal than vendor self-assessment, and a substantive validation for the company's plans to scale commercial driverless long-haul trucking operations. Source: Truck News source

Faraday Future (FF) Reports 242 Robot Units Shipped, March–June · humanoid ⚠️ target figure

FF reports that its robotics division shipped 242 units between March and June, ahead of expectations; Jia Yueting (FF's founder) has raised the full-year shipment target to 2,000 units accordingly. The 242 units represent actual deliveries; the 2,000-unit figure is a forward-looking target, not achieved capacity. Source: Sohu source

UBTECH Reports 5,000+ Pre-orders in 20 Days for New Humanoid Robot · humanoid ⚠️ pre-order figure

UBTECH (Chinese humanoid robot company) reports that its new life-like humanoid robot surpassed 5,000 pre-orders within 20 days, with the product formally unveiled the following day. Pre-orders are not deliveries; actual production capacity and fulfillment remain to be verified after the official launch. Source: MyDrivers source

Pudu Robotics Plans "Fully Robot-Staffed Hotel" with Shenzhen CITD · industrial ⚠️ planned

Pudu Robotics (Chinese service robot company) announced plans to build what it calls the world's first fully robot-operated hotel in Guangdong, China, targeting an opening in 2027. This is a planning-stage announcement, not a completed or operational deployment. Source: Let's Data Science source

IV. Industry Developments

Volkswagen Plans to Terminate Autonomous Driving Partnership with Bosch · autonomy

Volkswagen plans to exit its 2022 joint development agreement with Bosch on driver assistance and autonomous driving systems. After a cumulative investment of roughly €1.5 billion over four years, an internal assessment concluded the technology was "not yet competitive." Volkswagen intends to terminate the contract under its terms and pivot to directly procuring software and hardware from external suppliers, targeting a new engineering supply contract before September — a move that runs in parallel with a sweeping cost-reduction program that includes up to 100,000 job cuts and the closure of four factories in Germany. Notably, Volkswagen has taken a different approach in the Chinese market, pursuing local intelligent driving partnerships there, with the two regional strategies now visibly diverging. Source: Autohome source

South Korea Elevates Physical AI to National Strategy · adjacent

The South Korean government has released a comprehensive physical AI package: planned investment of 20 trillion KRW in manufacturing AI alongside a "National AI Data Repository" to preserve skilled craftspeople's expertise in digital form; a target to commercialize industry-specific humanoid robots across 10 sectors by 2028; and a goal to grow the robotics market from 5.6 trillion KRW in 2021 to over 20 trillion KRW by 2030, with one million robots deployed, and an ambition to become the global leader in physical AI by 2030. On the industry side, Korea Startup Forum and 11 companies jointly established a "Physical AI Industry Association," while Samsung and others announced major investments at sites including Gumi. The 2028/2030 targets are forward-looking government commitments, though the policy announcements and industry organization activity represent substantive real actions rather than single-vendor claims. Source: Seoul Economic Daily source

Xpeng Unveils Predictive World Model Framework X-Mind · world-model ⚠️ vendor claim

Xpeng (Chinese EV and autonomous driving company) presented its world model roadmap at the CVPR 2026 Embodied Foundation Model Deployment Workshop, releasing X-Mind — a predictive world model framework enabling vehicles to "simulate the future before deciding," building on the prior X-World, X-Foresight, and X-Cache lineage. The framework emphasizes proactive reasoning, controllable generation, and long-horizon prediction to support the next-generation VLA 2.0; a technical report is forthcoming. This is a capability roadmap announcement; production-level results remain to be validated. Source: XPENG source

Wujie Power Releases Latent-Space World Model MWA · world-model ⚠️ vendor claim

Following a $200M+ angel round last month, Wujie Power (Chinese embodied AI startup) has released MWA™, which it describes as the world's first "long-horizon bidirectional physical causal chain" latent-space world model. The company claims it abandons the next-frame video prediction paradigm in favor of modeling physical causality directly in latent space, and asserts it tops a leading embodied intelligence benchmark. "World's first" and "top-ranked" are vendor claims that require independent verification. Source: Jiqizhixin (Chinese AI media) source (WeChat, CN)

Zhiyuan Robotics (Chinese humanoid startup) Takes Stake in Force Sensor Developer Lanpoint Touch · hardware

Following its previously reported milestone of the 15,000th unit rolling off its production line, Zhiyuan Robotics has taken a stake in force sensor developer Lanpoint Touch (Chinese sensor startup), extending its supply chain upstream into six-axis force and tactile sensing to strengthen its standardized supply chain ecosystem. Source: Guandian.cn source

Deep Robotics Rebrands and Announces "1+X+N" Global Strategy · embodied

Hangzhou-based quadruped robot maker "Hangzhou Deep Robotics" has dropped the city name from its branding, rebranding simply as "Deep Robotics," and unveiled a "1+X+N" strategy at its Global Partner Summit to accelerate international expansion of commercial embodied intelligence. Source: Yahoo Finance source

Meta Reported to Acquire Robotics AI Company to Accelerate Humanoid Development · adjacent ⚠️ unconfirmed report

Reports indicate Meta has acquired a robotics AI company to accelerate its humanoid robot development; transaction details and the identity of the acquisition target have not been officially confirmed. Source: Stocktwits source

Hardware · Supply Chain

· Ruiyan Intelligent Control (Chinese dexterous hand supplier): Reported to be the exclusive dexterous hand supplier for dozens of Chinese embodied large-model companies, emerging as a hidden chokepoint as the dexterous hand segment scales (China's dexterous hand shipments are projected to grow from 19,200 units in 2025 to approximately 70,000 units in 2026). source (WeChat, CN)

· 3D Force-Tactile Sensors: A team from South China University of Technology has developed multiple dexterous hand designs incorporating a proprietary low-cost, high-sensitivity 3D force-tactile sensor, targeting cost reduction in manufacturing — tactile sensing is a key bottleneck in moving dexterous hands from "mechanical grasping" to "intelligent manipulation." source (WeChat, CN)

FutureX · Physical AI Daily — Issue 42 (06/29)

Shawn — Sun, 28 Jun 2026 15:35:31 +0000

Today's Highlights

· AgiBot (Chinese humanoid robotics company) rolled off its 15,000th embodied robot, less than three months after its 10,000th unit in late March — claiming a new global record in humanoid mass-production scale and speed. A 6-day factory livestream reported a task success rate of approximately 99.99% (livestream figures).

· Hyundai Motor Group will acquire SoftBank's remaining 9.7% stake in Boston Dynamics for $325 million, achieving full ownership. A production version of Atlas is planned to enter Hyundai's electric vehicle plant in Georgia, USA in 2028.

· Shanghai Innovation Institute and AgiBot released τ0-WM — the world's largest open-source embodied world model, with 5.5B parameters and approximately 27,300 hours of pretraining data, unifying future video prediction, action generation, and action evaluation into a single backbone.

· Tsinghua University open-sourced OpenHLM, a recipe for whole-body loco-manipulation VLA in humanoids: using less than half the demonstration data of strong baselines, it achieves an average task progress of 87.5% on long-horizon tasks (vs. 57.5% for GR00T N1.6).

· A UC San Diego team systematically analyzed world model "hallucinations," categorizing three failure types, showing these can be predicted using label-free signals (ρ≈0.8), and adapting models to new environments using just 50 trajectories — approaching expert-data performance.

I. Research Papers

τ0-WM: World's Largest Open-Source Embodied World Model · world-model

It unifies three previously separate capabilities — predicting future video, generating actions, and evaluating action quality — into a single video DiT backbone, allowing the world model to "imagine" candidate futures before execution and filter out actions that look plausible but fail to advance the task. Pretrained on 5.5B parameters and approximately 27,300 hours of heterogeneous data (including 17,800 hours of real-robot teleoperation), it has been validated on long-horizon dexterous manipulation tasks such as toolbox organization, zipping backpacks, and pipe coupling. Both the model and data are fully open-sourced, making it the largest open-source embodied world model to date.

Shanghai Innovation Institute × AgiBot · arXiv 2606.01027 https://arxiv.org/abs/2606.01027 · Analysis: 真的在读论文 source

OpenHLM: An Open-Source VLA Recipe for Humanoid Whole-Body Loco-Manipulation · locomotion

Addressing the view that "humanoids should be more than bipedal dual-arm platforms," Tsinghua researchers systematically answer how to build whole-body loco-manipulation: collecting data via joint-space whole-body teleoperation, retaining non-humanoid pretraining, generating absolute joint values via multi-step flow, and extending capability with low-cost data sources such as standing teleoperation and the HuMI framework (no real robot required). On the self-built HLM-12 long-horizon task suite, it achieves 87.5% average task progress using less than half the demonstrations of two strong baselines — far exceeding GR00T N1.6 (57.5%) and Ψ0 (48.8%), and approaching the whole-body teleoperation oracle (97.5%).

Tsinghua University Institute for Interdisciplinary Information Sciences (Gao Yang group) et al. · arXiv 2606.22174 https://arxiv.org/abs/2606.22174 · Analysis: 机器之心 source

World Model "Hallucinations" Can Be Predicted — and Fixed with Minimal Data · world-model

Generative world models often render visually smooth sequences that have already diverged from real physics — "silent errors" that are highly dangerous for systems relying on them for planning and control. The research categorizes hallucinations into three failure types: perceptual hallucination, action marginalization, and scene divergence. It then identifies three runtime signals requiring no ground-truth labels (Spearman ρ≈0.8 correlation with real error), arguing that hallucinations are fundamentally a data coverage problem. Using these signals for curiosity-driven data collection, just 50 trajectories suffice to adapt a model to a new environment and approach expert-data performance. The accompanying MMBench2 benchmark covers 427 hours, 210 tasks, and 10 domains.

Nicklas Hansen, Xiaolong Wang (UC San Diego) · Analysis: 具身智能漫谈 source

NavWM: A Unified World Model That Nearly Doubles Navigation Success Rate · autonomy

Most visual navigation policies operate in a "single deterministic future," causing oscillation and looping at complex junctions. NavWM uses a bidirectional Mamba backbone to unify perception, trajectory prediction, and future-frame generation — regressing multiple candidate paths in one pass via trajectory anchors, then using the world model to "visually simulate" each and select the best. At 1.5B parameters, it reduces absolute trajectory error by more than 30% compared to the next-best method, and achieves a zero-shot success rate of 0.44 on unseen scenes — nearly double that of competing approaches.

Institute of Automation, Chinese Academy of Sciences × Beihang University · Analysis: 集智实验室 source

LA4VLA: Decoupling Language-to-Action Pretraining from Vision · vla

In standard VLA training, language signals are overwhelmed by dense visual-action correlations — models appear to follow instructions but actually exploit visual shortcuts. Diagnostic experiments show that when vision and language conflict, models follow vision. Researchers from Shanghai Jiao Tong University and Alibaba address this by explicitly decoupling language-action pretraining: first learning how language instructions constrain continuous actions with no visual input, then combining this with standard VLA training — improving downstream policy performance and robustness under visual perturbation.

Shanghai Jiao Tong University (MINT) × Alibaba · github.com/MINT-SJTU/LA4VLA · Analysis: 具身智能之心 source

Other papers today: A survey on "World Action Models (WAM)" led by Nvidia researchers, mapping the rapidly emerging paradigm of "pretrain to imagine, fine-tune to act"; an ECCV 2026 benchmark of 303 questions revealing significant reasoning gaps in video generation models.

Open Source · Tools · Benchmarks

· Unitree Qmini: Unitree (Chinese robotics company) open-sourced a sub-$1,000 biped robot project with structural parts printable on consumer 3D printers, along with training code — a strategy of trading open-source access for data and ecosystem.

· Alibaba Qwen Language World Model: Alibaba open-sourced an agentic world model covering 7 environments including MCP, search, terminal, and software engineering, claiming to outperform GPT-5.4 on multiple metrics. Note: this is a language world model for agents, distinct from embodied/physical world models ⚠️ vendor figures.

· Zhicheng AI (Chinese physical AI startup) Chengling V0.1: Zhicheng AI open-sourced its "Chengling" physical intelligence world model V0.1 alongside a new funding round, and upgraded its TR4 Pro and TR5 Pro humanoid products.

II. Funding & Deals

Odyssey ｜ Series B ｜ $310 Million ｜ $1.45 Billion Valuation · world-model

Natural Capital led the round, with Amazon, GV, AMD Ventures, EQT, and the CIA's In-Q-Tel participating, along with a preferred cloud partnership with AWS. Founded by autonomous driving veterans, Odyssey focuses on interactive world models and "world simulation." This round makes it a unicorn and is another data point in sustained primary-market appetite for world model companies.

Source: 六观阿尔法 source

RoboScience ｜ Series A ｜ 1 Billion RMB · embodied

This company, founded in late 2024 and previously low-profile, made its public debut at an embodied large-model launch event, disclosing it had closed four funding rounds within a year, with Series A reaching 1 billion RMB. Founder Tian Ye studied under Andrew Ng and previously led Apple's AI platform; Chief Scientist Shao Lin is an NUS assistant professor and two-time ICRA Best Paper winner/nominee. Rather than scaling teleoperation, the company proposes VLOA (adding object trajectory "O" to VLA) combined with a proprietary physics engine, RoboMirage, to build an automated data flywheel — claiming to reduce per-sample data cost from several RMB to fractions of a cent. Cross-embodiment transfer rates and Sim-to-Real performance remain to be validated in real-world settings.

Source: 白鲸实验室 source

SnowOrigin ｜ New Round ｜ Backed by Gong Hongjia, Lu Qi, and Others · adjacent

Gong Hongjia (co-founder of Hikvision, Chinese surveillance giant), Lu Qi (founder of Qizhi Ventures, prominent Chinese tech investor), and overseas institutions have backed this neural interaction company. Using a neural wristband to capture surface electromyographic signals from the forearm, combined with first-person-view devices and AI, the company converts human hand poses and force dynamics into robot training data — targeting the human motion capture layer in the embodied AI data bottleneck.

Source: 高工人形机器人 source

LiberAI ｜ Pre-Series A ｜ Led by Shunwei Capital · world-model

Shunwei Capital (Chinese VC backed by Xiaomi's Lei Jun) led the round, with Cathay Capital, Yuanhe Origin, and Muhua Kechuang participating, and Sequoia China and Zhen Fund continuing to invest. Founded just six months ago with a team of under 30, the company focuses on physical world models and embodied intelligence, pursuing a "human UMI data + world model" approach with accompanying UMI hardware and data collection infrastructure.

Source: 高工人形机器人 source

SEAHI ｜ Series A ｜ Over 1 Billion RMB · adjacent

This round sets a global record for single-round funding in marine robotics. Founded by alumni of Harbin Engineering University, SEAHI specializes in underwater and marine robots — a signal that venture capital is expanding beyond land-based humanoids into underwater and special-purpose applications.

Source: Robot source

III. Commercial Deployment

AgiBot's 15,000th Robot Rolls Off the Line — Deployed Immediately · humanoid ⚠️ Livestream figures

On June 28, AgiBot announced the mass-production rollout of its 15,000th embodied robot (the Elf G2). According to the company, the 10,000th unit (Expedition A3) rolled off on March 30; scaling from 10,000 to 15,000 took under three months, compared to under four months to go from 5,000 to 10,000 — indicating accelerating ramp-up. AgiBot says it has established seven deployment categories: production line loading/unloading, industrial handling, logistics sorting, guided tours, retail, security patrol, and commercial cleaning, spanning full-size, half-size, wheeled humanoid, and quadruped platforms. Units go directly from the line into active deployment. A 6-day livestream connected to the earlier G2 factory broadcast reported approximately 99.99% task success — but this reflects livestream demonstration performance, not mass-production yield. Combined with the "Data Collection 2.0" system released on June 26 (covering 430+ real-world scenarios), the company's logic is to use mass-production scale to feed real-world data back into model and hardware improvement.

Source: 科创板日报 source、AgiBot https://en.prnasia.com/releases/apac/agibot-s-15-000th-robot-rolls-off-the-production-line-marking-a-new-milestone-in-embodied-ai-deployment-538948.shtml

Unitree G1 Powers America's First On-Demand Home Cleaning Service · humanoid ⚠️ Autonomy level unverified

US startup Gatsby has launched the first humanoid robot home cleaning service for general consumers in San Francisco, using Unitree's G1 robot at a flat rate of $150 per visit — below the local human cleaning rate of $150–$300. The company positions itself as a "robot-agnostic" service dispatch layer (Unitree provides the hardware; Gatsby handles scheduling and operations), treating real household task data as an asset to feed back into model improvement. Trial slots are fully booked, but only static images have been released and the proportion of remote human intervention has not been disclosed — leaving the robot's true level of autonomy unclear.

Source: 深观启元 source

Volkswagen ID. Buzz Robotaxi Begins Road Testing in Los Angeles · autonomy

Volkswagen's ID. Buzz autonomous vehicle fleet has begun road testing in Los Angeles in preparation for integration with the Uber platform, marking another move by a legacy automaker into the robotaxi space.

Source: Dimsum Daily https://www.dimsumdaily.hk/vw-id-buzz-robotaxi-begins-los-angeles-trials-ahead-of-uber-launch/

IV. Industry News

Hyundai Motor Acquires SoftBank's Stake for $325 Million, Taking Full Control of Boston Dynamics · humanoid

According to South Korean media, Hyundai Motor Group will acquire SoftBank's remaining 9.7% stake in Boston Dynamics for $325 million, achieving full ownership of the robotics company. SoftBank had retained a minority stake when it sold 80% to Hyundai in 2021 for approximately $880 million, and is now exiting via a contractual put option. Simultaneously, Hyundai plans to sell back to SoftBank its share of the RAI Institute (Robotics and AI Institute) — co-founded with Boston Dynamics in 2022 with cumulative investment of $424 million — for approximately $100 million, effectively splitting assets so that Hyundai retains the hardware company while SoftBank takes the advanced research institute. Hyundai had previously announced plans to deploy a production version of Atlas in its electric vehicle plant in Georgia, USA in 2028; full ownership gives it complete control at the critical juncture of humanoids moving from demonstration to commercial deployment.

Source: 高工人形机器人 source

Faraday Future Unveils Industrial Wheeled-Arm Robot "Faber" · industrial ⚠️ Vendor figures

Faraday Future (FF) unveiled the Faber industrial wheeled-arm robot and a six-series robot lineup at an event in Chicago, claiming entry into industrial automation. Given FF's prolonged track record of delivery delays and cash flow difficulties, the commercial viability of these products will need to be validated by future orders and deployments.

Source: 第一电动网 https://d1ev.com/newsflash/304762

China's First Embodied World Model "Wowu" Completes Generative AI Service Filing · world-model

Beijing Humanoid (National-Local Joint Embodied Intelligence Innovation Center), a state-backed Chinese robotics initiative, has completed China's generative AI service regulatory filing for its embodied world model "Wowu" — described as the first general humanoid robot foundation model to complete the full local compliance review process in China, providing a template for large-model-driven embodied products seeking regulatory clearance in the Chinese market.

Source: 观点网 https://www.guandian.cn/article/20260628/569754.html

Masayoshi Son Delays Retirement, Plans to Transform SoftBank into a "Robot Company" · adjacent ⚠️ Personal statement

Masayoshi Son has said he will delay retirement and continue for more than a decade, aiming to shift SoftBank's focus toward AI and robotics. Combined with SoftBank's move to acquire the RAI Institute from Boston Dynamics, SoftBank's bet on Physical AI is deepening.

Source: MSN https://www.msn.com/zh-cn/news/other/%E5%AD%99%E6%AD%A3%E4%B9%89%E6%8E%A8%E8%BF%9F%E9%80%80%E4%BC%91%E8%AE%A1%E5%88%92-%E6%8A%95%E8%BA%ABai%E9%A2%86%E5%9F%9F%E5%86%8D%E6%88%98%E5%8D%81%E4%BD%99%E5%B9%B4-%E5%BC%95%E9%A2%86%E8%BD%AF%E9%93%B6%E8%BD%AC%E5%9E%8B%E6%9C%BA%E5%99%A8%E4%BA%BA%E5%85%AC%E5%8F%B8/ar-AA26q5bh

onsemi Stock Drops ~24% After Synaptics Acquisition Announcement · hardware

Following its earlier announcement of an all-stock acquisition of edge AI chip maker Synaptics for approximately $7 billion — a move into Physical AI — onsemi shares fell approximately 24%, reflecting market concerns about share dilution and integration risk for what would be the company's largest-ever acquisition.

Source: TIKR.com https://www.tikr.com/blog/on-semiconductor-nasdaq-stock-hits-worst-day-since-march-2024-following-massive-all-stock-acquisition-of-synaptics

Hardware · Supply Chain

· Motors: Nearly 20 Chinese motor manufacturers are competing for humanoid robot joint and actuator orders, making upstream motors one of the most contested segments in the current humanoid supply chain.

· Wiring harnesses: A company in Ningbo, China produces over 1 billion hair-thin robot wiring harnesses annually, supplying the "vascular system" of humanoid robots — continuing the trend of scaled harness assembly seen with players such as Shizhi Zhihang and Tianhai Electronics.

· Reducers / bearings: Research reports have identified automotive component giants such as Schaeffler as key beneficiaries in the humanoid robot supply chain ⚠️ analyst report figures.

· Dexterous hands: AGILINK (spun off from AgiBot), reports cumulative delivery of over 8,000 OmniHand dexterous hands and over 10,000 grippers, claiming operating profitability in its first full quarter of operation ⚠️ unverified single-party figures.

FutureX · Physical AI Daily — Issue 41 (06/28)

Shawn — Sat, 27 Jun 2026 14:46:53 +0000

Today's Highlights

· USC and collaborators open-source Ψ₀, a VLA foundation model for whole-body loco-manipulation — using roughly 1/10 the data, it outperforms second-place GR00T N1.6 by over 40% on multiple long-horizon dexterous manipulation tasks.

· Sharpa (Chinese tactile AI startup), together with Tsinghua, Berkeley, and others, releases FTP-1, claiming it as the first general-purpose tactile foundation model, with a single policy spanning 21 heterogeneous tactile sensors and multiple embodiment types.

· Tashi Zhihang (Chinese robotics startup) signs an agreement with Tianhai Electronics (Chinese automotive wire harness leader) to advance large-scale robotic assembly of flexible wire harnesses, billing it as the world's first embodied-intelligence deployment on a wire harness production line.

· Xinjian Transmission (Chinese precision drive components maker), a key supplier of planetary roller screws for Tesla Optimus, has had its ChiNext IPO application accepted — robotics component makers are racing to list on China's A-share market.

· China's embodied intelligence sector raised over ¥46 billion in the first half of the year; global robotics investment hit a record of approximately $16 billion — even as debate over a "bubble squeeze" intensifies in parallel.

I. Research & Papers

Ψ₀: Open-source VLA Foundation for Whole-Body Loco-Manipulation, Emphasizing Data Efficiency · vla

Adding to the open-source humanoid foundation model lineup alongside GR00T and π0, Ψ₀ is notable for its data efficiency. Trained on roughly 800 hours of human video + 30 hours of real-robot data, it surpasses baselines trained on ten times the data across multiple long-horizon dexterous manipulation benchmarks, and outperforms second-place GR00T N1.6 by over 40%. Architecturally, it uses Qwen3-VL as the visual-semantic backbone with an MM-DiT flow matching action expert, and applies real-time chunking during training to mitigate action jitter caused by VLA inference latency. Weights and code are open-sourced.

USC PSI Lab × NVIDIA × WorldEngine · Commentary: AI 共享前沿 source (WeChat, CN)

FTP-1: Claimed First General Tactile Foundation Model, One Policy Across 21 Sensors · perception

Tactile sensing has long been trapped in a fragmented landscape — one sensor type, one hand, one task set per model — with GelSight, force/torque, and other sensors being mutually incompatible in input format, resolution, and form factor, making experience non-transferable. FTP-1 unifies 21 heterogeneous sensor types and multiple embodiment forms under a single policy, arguing that touch should have an independent, high-frequency, transferable representation and control pathway rather than serving merely as auxiliary input to a vision-language model. The team claims it as the first general-purpose tactile foundation policy.

Sharpa × Tsinghua University × UC Berkeley × Shanghai Jiao Tong University × ETH Zürich · Commentary: 新浪财经 source

Fei-Fei Li: A Functional Taxonomy of World Models — Breaking "World Model" into Three Types · world-model

Responding to the overuse of "world model" as a label for nearly every video generation system, Fei-Fei Li proposes a functional taxonomy: she divides world models by role into renderers (outputting pixels for human viewing), simulators (propagating state transitions according to physics/dynamics), and planners (selecting actions over trajectories) — the three forming a closed loop underpinning "spatial intelligence." The goal is to provide clearer vocabulary for design, evaluation, and toolchains, and to distinguish genuine world models from pure visual generation.

Fei-Fei Li (World Labs / Stanford) · Commentary: 梧桐区评论 source (WeChat, CN)

LeCun's Camp Adds Formal Theory to the JEPA World Model · world-model

Complementing Fei-Fei Li's functional taxonomy, Yann LeCun's team provides a mathematical grounding for JEPA (Joint Embedding Predictive Architecture): it defines the conditions under which JEPA can learn a faithful world model, and benchmarks show that current implementations remain quite fragile and fall well short of that standard. This provides theoretical and empirical support for LeCun's long-standing argument that large language models cannot serve as world models, and that world learning must happen in representation space rather than in word-level statistics.

Yann LeCun's team · Commentary: 图灵人工智能 source (WeChat, CN)

Other papers today: World Pilot — guiding a VLA with "world-action" priors to address the gap left by static image-text pretraining in continuous contact interactions source (WeChat, CN); GeoAlign — reinforcing geometric alignment in VLAs to improve precision tasks such as gripper edge-contact and transparent object localization source (WeChat, CN); Looped World Models — a 1B-parameter recurrent world model claiming substantial parameter efficiency gains over baselines ten times its size source (WeChat, CN); World Action Models as zero-shot policies, jointly modeling perception, action, and future states with a shared Transformer source (WeChat, CN); Berkeley "Do as I Do" — teaching robots to self-learn multiple dexterous manipulation skills by watching YouTube videos source (WeChat, CN); a single-line optimizer change that mitigates degradation of language instruction following after VLA fine-tuning source (WeChat, CN).

Open Source · Tools · Benchmarks

· Qwen-AgentWorld / Qwen-RobotWorld: Alibaba is progressively open-sourcing three categories of "world model" frameworks targeting language, virtual, and physical worlds, using community engagement to build research influence source (WeChat, CN)source (WeChat, CN).

· LabVLA: Zhejiang University's zjunlp lab releases an embodied VLA model for scientific laboratory scenarios, with weights and code open-sourced on HuggingFace and GitHub — pushing AI for Science from "thinking" to "doing" source (WeChat, CN).

· Kaiwu World Model 3.0: Released by Daxiao Robotics (Chinese embodied AI startup), a subsidiary of SenseTime (Chinese AI company), which claims it as the first open-source world model in China to achieve commercial deployment source (WeChat, CN).

II. Funding & Deals

Simple AI (深朴智能) ｜ New Round ｜ Hundreds of millions of RMB · embodied

Linear Capital and Puhua Capital led the round, with Junshan and others following; this is the third round of hundreds-of-millions-of-RMB funding within six months for this general-purpose embodied intelligence robotics company, which focuses on training embodied foundation models via a "human learning" approach. Proceeds are earmarked for model R&D and multi-scenario validation. Source: 百数朝智网 source (WeChat, CN)

10Beauty ｜ New Round ｜ $23.5 million (~¥160 million) · adjacent

The Boston-based AI nail care robotics company will use the funds to deploy 850 fully automated nail service units and expand into retail and hotel channels. A former DJI executive has joined this niche service robotics segment, illustrating how embodied AI deployment is penetrating high-frequency, standardized vertical consumer scenarios. Source: 具身涌现 source (WeChat, CN)

FaceMind (脸谱心智) ｜ Pre-A ｜ Tens of millions of RMB · world-model

Xinglian Capital joined as a new investor while existing backer 360 (Chinese internet conglomerate) over-subscribed; FA firm Shendu Capital indicated a new round is already in progress. The company entered the world model space via anime/virtual content, making it one of the early-stage players to raise capital amid the current world model wave. Source: 融了么 source (WeChat, CN)

Innok Robotics (Germany) ｜ New Round ｜ Over €3.3 million · industrial

Headquartered near Regensburg, Innok specializes in mobile robots for outdoor and brownfield (legacy facility) environments; the funds will accelerate business expansion. This is a relatively rare early-stage European robotics fundraise in the current news cycle. Source: 阿西莫夫的机器人 source (WeChat, CN)

Lanjian Relu ｜ New Round ｜ Amount undisclosed · industrial

Founded less than two years ago, this Chinese robotics company targets construction sites and outdoor heavy-duty environments, positioning itself as "the next generation of labor." Beyond humanoids — which struggle with real-world deployment — construction and engineering operations represent one of the few sectors with demonstrably clear commercial demand. Source: 科创界 TechIn source (WeChat, CN)

Daxiao Robotics (SenseTime subsidiary) ｜ Angel+ Round ｜ Amount undisclosed · world-model

SenseTime's embodied intelligence spinout, which has released the Kaiwu World Model 3.0, has secured an Angel+ round following cumulative fundraising of several hundred million dollars in the first half of the year — continuing the trend of major tech-company embodied AI units spinning out and raising independent capital. Source: 企投通 source (WeChat, CN)

III. Commercial Deployment

Galaxy General (银河通用) Reports Order Surge for Galbot G1, Rushes to Expand Production · embodied ⚠️ Unverified — single-source claim

Following its Spring Festival Gala appearance, Galaxy General (Chinese humanoid robotics startup) says orders for its Galbot G1 have surged and the company is urgently ramping production to accelerate deliveries. Building on its earlier ¥236 million contract for 500 units from Yibin city and the Galbot S1's deployment in a CATL battery cell factory, the company's "showcase → order → delivery" cycle appears to be tightening. However, the reported order volume is from a single secondary source; specific unit counts and customer identities have not been independently verified. Source: MSN repost source

Tashi Zhihang Partners with Tianhai Electronics to Scale Robotic Assembly of Automotive Flexible Wire Harnesses · embodied ⚠️ Partnership announcement

The two companies have signed a strategic cooperation agreement to jointly advance intelligent upgrading of automotive wire harness assembly, claiming this will be the world's first large-scale embodied intelligence deployment on a wire harness production line. Tashi Zhihang (Chinese dexterous manipulation startup) combines multimodal perception, force control, and its general-purpose embodied foundation model AWE3.0; its A1 robot previously set a Guinness World Record for completing over 100 sub-millimeter (0.5mm²) wire harness assembly insertions within one hour. At this stage, this is a cooperation agreement — "large-scale" remains a target rather than an existing production capability. Source: TOM 科技 source

ROBOTIS to Build Humanoid Robot Factory in Tashkent · humanoid

South Korean robotics company ROBOTIS plans to build a humanoid robot factory in Tashkent, Uzbekistan, extending its production capacity and supply chain into Central Asia. This is a capacity-building move; production timelines and scale have yet to be disclosed. Source: Zamin.uz source

Jiyi Intelligent Debuts New Products, Robots Head to the Factory Floor · industrial ⚠️ Launch event claims

Jiyi Intelligent (Chinese industrial robotics startup), founded about 8 months ago, held a launch event and unveiled two complete robot models focused on operating directly in factory environments. As a new company's debut, the factory demonstrations are more a capability showcase than evidence of scaled production-line deployment; real-world depth remains to be validated. Source: 新华网 source

Caocao Mobility Brings Robotaxi Service to Abu Dhabi · autonomy ⚠️ Expansion announcement

Caocao Mobility (Chinese ride-hailing platform, Geely affiliate) is launching its Robotaxi service in Abu Dhabi, seen as an entry point for Geely's overseas plan involving 100,000 autonomous vehicles; Caocao contributes operations capability while Geely provides customized vehicles and supply chain support. This is currently a market entry and operations setup phase; scaled operational data have not yet been provided. Source: 看中东 DeepMENA source (WeChat, CN)

IV. Industry Developments

UBTECH Launches Wheeled Industrial Humanoid Robot Cruzr Y1 · industrial ⚠️ Launch event claims

The Cruzr Y1 made its global debut at an OPPO supply chain technology exhibition, targeting flexible operations including bin/carton destacking, palletizing, machine tending, and sorting — addressing the pain points of traditional automation such as halting on minor material misalignment and slow changeover. Hardware includes Chinese-made Horizon Robotics S100P/S600 (Chinese AI chip maker) chips, 360° omnidirectional wheels with a vertical lift module, and integrated harmonic joint modules; fully loaded runtime exceeds 4 hours with hot-swap battery support for 24-hour shift operation, and a full SDK is open for secondary development. This is a new product demonstration; mass production and deployment figures have not been released. Source: 名企参访直通车 source (WeChat, CN)

XPENG Robotics Sprints Toward Mass Production; He Xiaopeng Takes Personal Charge as Product Head · humanoid

XPENG has restructured its robotics business: nine new departments have been established, with chairman He Xiaopeng personally taking on the role of head of the product division — aligned with the target of mass-producing the IRON humanoid by year-end and the company's global expansion plans. The move is widely interpreted as He Xiaopeng treating robotics as a make-or-break bet, with the organizational overhaul designed to capture the mass-production window. Source: 界面新闻 source

Xinjian Transmission's ChiNext IPO Accepted; Key Tesla Optimus Screw Supplier Eyes A-Share Listing · hardware

The IPO application of Xinjian Transmission — active in dexterous hand components and screw assemblies — has been accepted, with proceeds targeting an "annual production of 1 million humanoid robot and automotive planetary roller screw" industrialization project. The company is identified as a core supplier of planetary roller screws for Tesla's Optimus. On the supply chain side, "picks-and-shovels" players such as Haiqing Zhiyuan and Xianggong Intelligence are also successively entering the capital markets, reflecting a flow of capital from complete robot systems upstream toward components with more predictable demand. Source: 猎场备忘录 source (WeChat, CN)

Zhiyuan Robotics Partners with Haojing Technology, Targeting Telecom Operator Market · embodied

Zhiyuan Robotics (Chinese humanoid startup) is partnering with Haojing Technology (Chinese telecom software provider) to bring its product portfolio — the Yuanzheng A2 (reception and guidance), Spirit G2 (continuous industrial operation), and Kutuo D1 (patrol and mobility) — into telecom operator channels, with operator retail outlets becoming a new distribution entry point for embodied AI deployment. Source: 名企参访直通车 source (WeChat, CN)

H1 2026: China's Embodied AI Funding Tops ¥46 Billion; Global Robotics Investment Hits Record; "Bubble" Debate Heats Up · adjacent

In the first half of 2026, disclosed funding for China's embodied intelligence and robotics sector exceeded ¥46 billion across 288 deals; global robotics investment also reached a record high of approximately $16 billion. At the same time, discussion of "the bubble is deflating — the squeeze has already begun" is growing louder within and outside the industry. Hot money flowing in and valuations needing to be absorbed are the defining feature of the sector right now. Source: 于仁颇黎 source (WeChat, CN)｜HOKANEWS source

China's SAMR: Accelerate Standards Development for AI Agents, Embodied Intelligence, and Other Frontier Fields · adjacent

China's State Administration for Market Regulation (SAMR) has stated it will accelerate the pace of standards development for frontier technology areas including AI agents and embodied intelligence. Layered alongside the recent submission of L3/L4 autonomous driving national standards for approval and the adoption of a UN global technical regulation on autonomous driving systems, the standards and regulatory "infrastructure" is being built out rapidly. Source: 新浪财经 source

Momenta's Hong Kong IPO Said to Have Secured Mercedes-Benz, GIC as Cornerstone Investors · autonomy ⚠️ Unverified — sources claim

After clearing the Hong Kong Stock Exchange hearing and racing to be the "first Physical AI public company," sources indicate that Mercedes-Benz and GIC (Singapore's sovereign wealth fund) will participate as cornerstone investors in Momenta's Hong Kong IPO. If confirmed, this would strengthen the capital backing for its autonomous driving mass-production-to-Robotaxi data flywheel; the information comes from unnamed sources and has not been officially confirmed by the company. Source: The Business Times source

Hardware · Supply Chain

· Core Components (Japanese Media Perspective): Japanese media report that Japan has fallen significantly behind China in humanoid robot complete systems, but may still have an opportunity to catch up in core "bones and muscles" components such as reducers, servo motors, and torque sensors source (WeChat, CN).

· Linkerbot (灵心巧手): Claims over 80% global market share in high-degree-of-freedom dexterous hands, with monthly production capacity of 4,000 units, cumulative deliveries exceeding 10,000 units, and a full-year 2026 delivery target of 50,000–100,000 units (vendor-stated figures; ⚠️ unverified) source (WeChat, CN).

· Screw Demand in Dexterous Hands: As active degrees of freedom migrate into the forearm, a single dexterous hand now requires close to 12 active DOF — meaning 12 or more screws per hand — a positive demand signal for upstream screw manufacturers source (WeChat, CN).

FutureX · Physical AI Daily — Issue 40 (06/27)

Shawn — Fri, 26 Jun 2026 14:49:49 +0000

Today's Highlights

· onsemi is proposing to acquire edge AI chip maker Synaptics in an all-stock deal worth approximately $7 billion, entering the physical AI space in what would be the company's largest acquisition to date.

· Chinese embodied AI startup Wujie Power (Wujie Dongli) closed an angel round exceeding $200 million led by JD.com, bringing cumulative angel-stage funding to approximately RMB 2.8 billion within six months; the company also claims nearly $100 million in orders.

· General Intuition's $320 million Series A is finalized at a $2.3 billion valuation — led by Khosla Ventures following mid-June reports of negotiations, with participation from Bezos, Schmidt, and others.

· The largest open-source teleoperation dataset to date, ABC-130K (3,500 hours, 130,000 episodes, 195 tasks), has been released along with the full hardware, training, and simulation stack.

· Chinese physical AI foundation model startup Shendu Jizhi has raised another round of several hundred million RMB, focusing on training embodied foundation models using a "human learning" approach.

I. Research

Largest Open-Source Teleoperation Dataset ABC: 3,500 Hours, 130,000 Trajectories · benchmark

A major addition to open-source infrastructure for imitation learning: ABC releases the largest teleoperation dataset to date — ABC-130K — comprising 3,500 hours, over 130,000 episodes, and 195 tasks, together with a low-cost hardware design, training framework, and simulation pipeline, providing a fully reproducible full-stack starting point for behavior cloning.

Arthur Allshire et al. · arXiv 2606.27375 source

World Model "Hallucinations" Are Predictable — and Preventable · world-model

Generative world models are becoming increasingly realistic, yet often silently drift from true dynamics while producing smooth-looking video. Nicklas Hansen and Xiaolong Wang show that such hallucinations are concentrated in low-coverage regions of the state-action space and can be localized and suppressed using lightweight data-side signals, offering a diagnostic approach for reliable world model rollouts.

Nicklas Hansen, Xiaolong Wang (UC San Diego) · arXiv 2606.27326 source

PhysiFormer: Learning Physical Simulation in World Coordinates · world-model

Most video world models predict future frames in viewpoint-dependent pixel space, often producing physically inconsistent results. Andrea Vedaldi et al. propose PhysiFormer, which uses a diffusion transformer to represent objects as 3D meshes in world coordinates: given initial vertex positions, velocities, and material properties (rigid/elastic), it samples future vertex trajectories — a step toward controllable, reusable physical simulation.

Yiming Chen et al. · arXiv 2606.27364 source

Play2Perfect: Dexterous "Play" Pretraining for Precision Assembly · manipulation

Precision assembly is both contact-dense and reward-sparse — difficult for imitation learning to collect data on and for RL to explore. Stanford's C. Karen Liu, Jeannette Bohg, and colleagues systematically analyze what is actually useful in multi-finger "play"-style pretraining, yielding a transferable pretraining recipe for the long-standing challenge of high-precision assembly.

Tyler Ga Wei Lum et al. (Stanford) · arXiv 2606.26428 source

CoStream: Composing Simple Behaviors for Generalizable Complex Manipulation · manipulation

Long-horizon, contact-dense tasks such as inserting a GPU into a PCIe slot require both millimeter-level precision and zero-shot generalization to new tasks. Stanford's Wenlong Huang et al. propose CoStream, which generates complex manipulation by composing simple behaviors, easing the generalization bottleneck of end-to-end policies.

Haonan Chen et al. (Stanford) · arXiv 2606.26423 source

First Generalization Theory for JEPA World Models · world-model

JEPA (Joint Embedding Predictive Architecture), championed by LeCun as a viable path to world models, has demonstrated empirical success without theoretical underpinning. Wang Yisen et al. at Peking University present the first generalization theory for JEPA world models, explaining why predicting dynamics in latent space rather than reconstructing pixels works, providing theoretical grounding for this approach.

Jingyi Cui et al. (Peking University) · arXiv 2606.27014 source

OmniContact: Humanoid Loco-Manipulation via Contact Flow Chaining of Primitive Skills · locomotion

Long-horizon humanoid loco-manipulation requires both robust execution of individual primitive skills and seamless, closed-loop, self-recoverable chaining between them. OmniContact uses a contact flow to chain primitive skills, improving generalization and robustness for humanoid loco-manipulation.

Runyi Yu et al. · arXiv 2606.26201 source

VLA+RL Approach Wins ICRA 2026 LeHome Challenge in Laundry Folding · manipulation

Bimanual laundry folding is a notoriously difficult task given dense contacts and large deformations. This approach won first place in the simulation track among 62 teams at ICRA 2026 LeHome Challenge and runner-up in the real-robot track, by adding a reinforcement learning closed loop to a VLA policy.

Ilia Larchenko · arXiv 2606.27163 source

Other papers today: TaskNPoint — teaching a humanoid to hit a tennis backhand in minutes (2606.26215); "Ace" 8-DOF table tennis robot hardware design targeting professional-level play (2606.26643); Bridging Performance and Generalization — preventing agile flight RL policies from failing upon environment transfer (Scaramuzza group, 2606.27348); Fast LeWorldModel — accelerating JEPA visual planning (2606.26217); Einstein World Models — using counterfactual imagination to extend complex reasoning beyond language (2606.26969); VibeAct — driving contact-dense dexterous manipulation using vibration signals from piezoelectric microphones (2606.27344); HumanoidUMI — driving whole-body humanoid manipulation from robot-free demonstration data (2606.27239).

Open Source · Tools · Benchmarks

· PhysEditWorld: A large-scale dataset for "physically editable" world models that treats physical dynamics as explicit controllable variables rather than implicit correlations (arXiv 2606.26694 source).

· NavIsaacLab: An open-source benchmark for crowd-aware navigation that generates realistic pedestrian crowds in parallel on Isaac Sim (arXiv 2606.26265 source).

· OctoSense: An open-source multimodal sensing platform featuring stereo RGB + event cameras, LiDAR, thermal imaging, IMU, and RTK positioning (UPenn, Kostas Daniilidis et al., arXiv 2606.27317 source).

· ForesightSafety-VLA: A unified safety diagnostic benchmark for VLA models (arXiv 2606.27079 source).

II. Funding & Deals

Wujie Power (Wujie Dongli, Chinese general-purpose embodied AI startup) ｜ Angel Round ｜ Over $200 million (~RMB 1.36 billion) · embodied

Led by a JD.com-affiliated fund, with participation from C Capital, Hony Capital, Shengyu Investment, Fengyuan Investment, and continued backing from existing investors Linear Capital, Sequoia China, Huaye Tiancheng, and Yarui Capital. Founder Zhang Yufeng, formerly head of intelligent driving at Horizon Robotics (Chinese automotive AI chip company), leads the company's general-purpose embodied "large-brain small-brain integrated" approach. The second-generation wheeled dual-arm robot K15 has entered mass production with 1,200 TOPS of compute. Including prior rounds, cumulative angel-stage funding totals approximately RMB 2.8 billion; the company also claims nearly $100 million in global orders (per company disclosure). Source: Caixin source

General Intuition ｜ Series A ｜ $320 million ｜ Valuation $2.3 billion · world-model

Following reports of negotiations in mid-June, the round was led by Khosla Ventures, with participation from Bezos, Schmidt, and others; total disclosed funding reaches approximately $454 million. The company's approach is to train spatiotemporal reasoning capabilities using action labels from gameplay recordings — specifically which buttons were pressed and when — treating large volumes of game replays as training data for AI agents learning to act in the real world. Source: TechCrunch source

Shendu Jizhi (Chinese physical AI foundation model startup) ｜ New Round ｜ Several Hundred Million RMB · world-model

Following a prior round also in the hundreds of millions of RMB, the company has raised another several hundred million RMB to advance physical AI foundation model development. Its approach centers on a "human learning" paradigm, positioning itself against overseas players who bet on human demonstration data, and aiming to train embodied foundation models in a manner more closely modeled on how humans learn. Source: Sina Finance source

Zhicheng AI ｜ Two Rounds in Two Months · world-model

Led by a former CTO of Huawei's U.S. research lab, the company closed two funding rounds within two months. During this period it open-sourced its "Chengling" physical intelligence world model V0.1 and launched updated humanoid products TR4 Pro and TR5 Pro. Source: Zhidongxi source

Oula Wanxiang ｜ Angel Round ｜ Several Hundred Million RMB · embodied

Invested by China Merchants Capital (investment arm of state-owned conglomerate China Merchants Group), with funds directed toward developing an in-house embodied world model, scaling robot hardware manufacturing, and building a global developer ecosystem. Source: Central Enterprise Investment Association source (WeChat, CN)

Juwei Technology ｜ Series B1 ｜ Several Hundred Million RMB ｜ Valuation Nearly RMB 4 Billion · embodied

Closed a Series B1 round in the hundreds of millions of RMB, with a post-money valuation of nearly RMB 4 billion. Source: Guandian.cn source

Qunqing Intelligence ｜ New Round ｜ Nearly RMB 100 Million · industrial

The company focuses on embodied intelligence that "grows from the factory floor," targeting real-world industrial deployment and closed-loop operation. Source: 36Kr Hardtech source (WeChat, CN)

III. Commercialization & Deployment

UBTECH Claims ~5,000 Companion Robot Pre-orders in Two Weeks · humanoid ⚠️ Pre-order figures

UBTECH (Chinese humanoid robot company) reports nearly 5,000 pre-orders for its consumer companion robot (U1) within two weeks, attempting to break into the consumer market. Pre-order numbers remain separate from actual deliveries and retention; whether this converts to sustained shipments remains to be seen. Source: Pandaily source

Waymo Registers in Germany, Continues Expanding Nashville Fleet · autonomy

Waymo has completed legal entity registration in Germany, laying groundwork for European market entry, while continuing to expand its fleet in Nashville, U.S. Registration and road testing remain early-stage steps; large-scale commercial operations are still some distance away. Source: Automotive World source

Cao Cao Mobility Partners with K2 to Deploy First Robotaxis in Abu Dhabi This Year · autonomy ⚠️ Planned figures

Cao Cao Mobility (Chinese ride-hailing platform, Geely subsidiary) has signed a strategic partnership with Abu Dhabi's K2 Group, with plans to deploy its first Robotaxis within the year, continuing China's autonomous driving push into overseas markets. Source: China Daily source

BMW Spartanburg Figure 03 Logistics Deployment (Previously Reported) · humanoid

The Figure 03 deployment at BMW's Spartanburg plant, reported last week, received follow-up coverage this week from autoevolution and other international outlets. No new developments. Source: autoevolution source

IV. Industry Developments

onsemi to Acquire Synaptics for ~$7 Billion, Betting on Physical AI · hardware

U.S. analog chipmaker onsemi has announced an all-stock acquisition of touch and edge AI chip maker Synaptics for approximately $7 billion, at an exchange ratio of 1.350 shares representing a ~19% premium, with the deal expected to close in mid-2027 — onsemi's largest acquisition to date. onsemi aims to extend its capabilities from power and sensing into "intelligent systems," using Synaptics' edge AI compute, human-machine interface, and wireless connectivity to strengthen its physical AI portfolio; the combined company claims it could expand its total addressable market by approximately $30 billion to $243 billion by 2030. onsemi shares fell on the news, with the market skeptical about the valuation and integration outlook. The deal illustrates how competition in physical AI is extending upstream from robot hardware into underlying compute and sensing chips. Source: SiliconANGLE source

UBTECH Launches Wheeled Industrial Humanoid Cruzr Y1 · humanoid ⚠️ Launch/demo figures

UBTECH debuted its wheeled industrial humanoid Cruzr Y1 at an OPPO supply chain technology expo, demonstrating warehouse operations including bin destacking and restacking for factory logistics applications. A trade show demonstration remains distinct from stable mass production and large-scale delivery. Source: Guandian.cn source

Zhiyuan Robotics Releases Embodied Data Collection 2.0, Claims First National Embodied Evaluation Platform · adjacent ⚠️ Vendor claims

Zhiyuan Robotics (Chinese humanoid robot startup) has launched its embodied intelligence "Data Collection 2.0" technology framework and claims to have built China's first embodied intelligence robot evaluation platform, focusing on data acquisition and evaluation infrastructure. Source: Sina Finance source

XPENG Robotics Restructures, He Xiaopeng Takes Direct Charge of Product Division · humanoid

XPENG Robotics has reportedly reorganized into nine divisions, with He Xiaopeng (XPENG CEO) directly heading the product division to accelerate mass production. XPENG also states its second-generation VLA has received global regulatory approval alongside the UN ADS GTR (the regulation itself was previously reported) and plans international launches in 2027. Source: Sina source

LG Electronics and Nvidia Deepen Physical AI Partnership · adjacent

LG Electronics and Nvidia have further deepened their physical AI collaboration to co-build a physical AI ecosystem. The partnership is at a strategic level; specific milestones remain to be seen. Source: Seoul Economic Daily source

Renault Introduces Next-Generation Factory Robot Calvin · industrial

Renault has unveiled its next-generation factory robot Calvin, aimed at automating operations within its manufacturing network. Source: renaultgroup.com source

Hardware · Supply Chain

· Chinese key component localization: Domestic production of three critical components — harmonic drives, ball screws, and six-axis force/torque sensors — is accelerating in China, widely seen in the industry as key to driving down the cost curve for humanoid robots (industry view; specific cost reduction figures not yet confirmed).

· Li Auto develops in-house "Mach M100" chip: Li Auto (Chinese EV maker) has showcased its in-house intelligent driving chip, the Mach M100, a further step toward self-sufficient on-vehicle compute.

FutureX · Physical AI Daily — Issue 39 (06/26)

Shawn — Thu, 25 Jun 2026 14:51:18 +0000

Today's Highlights

· The UN's Global Technical Regulation on Automated Driving Systems (co-led by China) was approved at Geneva — the world's first such unified regulation; China's mandatory national standards for L3/L4 have been finalized and are now in the official approval process.

· BMW has launched deployment of Figure 03 humanoid robots for logistics sequencing at its Spartanburg plant, building on an eleven-month Figure 02 pilot in the body shop that assisted in producing over 30,000 X3s.

· Galaxy General Robotics (Chinese humanoid startup) Galbot S1 heavy-load humanoid has entered a CATL battery cell factory, achieving routine production operation in a new-energy manufacturing setting.

· Unitree Robotics (Chinese robot maker) has cut the retail price of its R1 humanoid robot to ¥29,900 with immediate stock availability and no waitlist, marking another step down in China's consumer-grade humanoid price war.

· Commercial service robot leader Pudu Robotics (Chinese service robot company) has closed a new funding round of nearly ¥1 billion, pushing its post-money valuation above ¥10 billion.

I. Research Papers

Causal-rCM: Distilling Autoregressive Video Diffusion into Real-Time Streaming · world-model

Action-conditioned interactive world models have long been bottlenecked by the multi-step sampling cost of video diffusion. This work extends the advanced diffusion distillation framework rCM to autoregressive video diffusion, unifying teacher-forcing and self-forcing training paradigms with an open-source recipe, providing a faster generation path for real-time streaming video and interactive world models. It ranked among the top papers on Hugging Face on the day of release.

Kaiwen Zheng et al. (Tsinghua University) · arXiv 2606.25473 source

In-Context World Modeling: Enabling VLAs to Generalize Across Settings via Contextual World Models · world-model

Most vision-language-action models (VLAs) condition only on the current observation and instruction, failing when the camera viewpoint or robot morphology changes. This method introduces in-context world modeling into the policy, using a small number of interaction demonstrations to infer the dynamics of new settings online, maintaining usability on unseen viewpoints and heterogeneous embodiments — targeting VLAs' widely criticized tendency to fail when the setup changes.

Siyin Wang et al. · arXiv 2606.26025 source

ForceBand: Teaching Robots Contact Force via EMG Signals · manipulation

Human teleoperation is a scalable data source, but motion capture trajectories and internet videos capture only motion and appearance, discarding the contact forces that are critical for force-sensitive manipulation. Jitendra Malik's team uses wearable surface electromyography (sEMG) to simultaneously record muscle effort during action collection, teaching policies "how hard to push," providing a new human-data channel for force-sensitive tasks such as twisting, pressing, and inserting.

Botao He et al. · arXiv 2606.26093 source

Memory Retrieval: General-Purpose Memory Retrieval for Long-Horizon Visuomotor Policies · manipulation

In partially observable environments such as homes, robots need to remember where objects were placed, what step a partner has completed, or when an appliance was switched on. This work proposes a general memory retrieval mechanism that lets visuomotor policies retrieve relevant information from past experience on demand to support long-horizon autonomy, rather than relying on task-specific memory modules.

Rutav Shah et al. (UT Austin) · arXiv 2606.25136 source

Reflective VLA: Incorporating Action Consequences into Context to Improve Generalization · vla

Most VLAs are reactive — predicting the next action from the current instruction and observation alone, implicitly assuming the current observation fully determines the action. This work has the model explicitly incorporate the causal consequences of actions into its context; with this reflection mechanism, VLA generalization and error correction in out-of-distribution scenarios improves.

Qing Lian et al. · arXiv 2606.25215 source

Learning Action Priors: Pre-Training Action Modules for Cross-Embodiment Manipulation · vla

Most VLAs attach an action head to a VLM backbone and train jointly — the vision-language priors are strong, but the action module learns physical motion nearly from scratch. This work learns action priors for the action module separately before cross-embodiment transfer, alleviating data scarcity on the action side, and reports approximately 68% success rate across 13 cross-embodiment tasks.

Dong Jing et al. · arXiv 2606.26095 source · Commentary: Embodied Intelligence Notes source (WeChat, CN)

WOLF-VLA: Extending VLAs to Whole-Body, Contact-Rich Humanoid Motion · locomotion

VLAs have demonstrated strong generalization in robotic arm manipulation, but whole-body, contact-rich humanoid motion remains largely unexplored due to scarce data and a lack of dynamically consistent demonstrations. The DFKI team proposes a VLA framework for humanoid whole-body optimal motion, encoding optimality and safety constraints into the learning pipeline to address this gap.

Melya Boukheddimi et al. (DFKI) · arXiv 2606.25591 source

Other papers today: FORCE (value-calibrated warm-up + self-distillation, improving sample efficiency for VLA reinforcement fine-tuning and breaking the imitation ceiling); GRAFT (part-correspondence-based schematic affordance transfer for few-shot grasping generalization); DeformGen (dynamics-based topological augmentation for deformable object manipulation demonstration expansion); StairMaster (three-stage RL for quadruped climbing high-risk open-riser stairs); RGB (RL-guided whole-body MPPI humanoid control); UniTeD (unified temporal diffusion for joint perception and planning in end-to-end autonomous driving); RAVEN (spatiotemporal memory for long-horizon navigation question answering).

Open Source · Tools · Benchmarks

· TacVerse: A multi-sensor dataset and benchmark for cross-sensor visuotactile perception, evaluating generalization across different tactile sensor designs (arXiv 2606.25877 source).

· 1000 Rallies: An event-camera dataset for robot table tennis plus real-time learned ball state estimation, targeting high-speed, low-latency perception (arXiv 2606.25620 source).

· 3D-Printable Tactile Test Set: Unified 3D-printable texture samples for fair cross-sensor comparison of tactile sensors, improving reproducibility (arXiv 2606.25886 source).

· Realsource: An open-source dataset release for robot training (source: All-About-Industries source).

· Alibaba's Qwen (Chinese AI lab) open-sources an agent-oriented language world model aimed at reducing agent errors; the team claims a 3B-activation model matches leading large models ⚠️ vendor claim (source: Qiling Model Plaza source (WeChat, CN)).

II. Funding & Deals

Pudu Robotics | New Round | ~¥1 Billion | Post-Money Valuation Exceeds ¥10 Billion · adjacent

Longgang Financial Holdings and Asia Investment Capital co-led, with participation from BAIC Industrial Investment, Lens Technology, Shanghai Honghui, and government-guided funds from the Pearl River Delta and Yangtze River Delta regions. Pudu Robotics focuses on commercial service robots (delivery/cleaning) and is extending into embodied AI under a "one brain, multiple forms" approach. According to Frost & Sullivan, it holds a 23% share of the global commercial service robot market, has shipped over 120,000 units, grew revenue more than 100% year-over-year in 2025, and is approaching EBITDA breakeven, with cumulative funding now exceeding ¥2 billion.Source: Hangfan Robotics source (WeChat, CN)

Senad | Series C | ¥300 Million · industrial

Investors include Full Truck Alliance, Haitong Kaiyuan, Oriza Prunus Ventures, Huayi Venture Capital, and Oriza Houwang, with existing investors Dachen Caizhì, Changshi Capital, and Chuhui Capital participating at above-quota allocations. Senad (Chinese logistics robot startup) focuses on loading/unloading and logistics embodied robots, targeting heavy manual tasks such as cargo handling in warehousing and long-haul logistics.Source: Tech Capital Circle source (WeChat, CN)

Luoshi Robotics (Chinese robot arm maker) | Hong Kong IPO (Listing Hearing Passed) · industrial

Luoshi Robotics has passed its Hong Kong Stock Exchange listing hearing, targeting the title of "first full-range intelligent robotics stock" on the Hong Kong market, with CICC and Guotai Junan as joint sponsors. Shareholders include the National Manufacturing Transformation and Upgrade Fund, New Hope Group, and Plum Ventures; its robotic arm business accounts for nearly 70% of revenue.Source: Sina Finance source

Hong Kong Robot IPO Pipeline | Ruiwei Technology, Eacon Mining Technology · autonomy

Concurrently, Ruiwei Technology (Chinese vision-embodied AI firm) has also passed its Hong Kong listing hearing, aiming for the title of "first visual embodied intelligence stock"; Eacon Mining Technology (Chinese autonomous mining truck company, backed by CATL investment and claiming the largest autonomous mining truck fleet globally) has likewise passed its Hong Kong listing hearing. Chinese robotics companies are clustering their Hong Kong listings in mid-2026, with exit channels opening simultaneously across the sector.Source: Sohu source · thebambooworks source

Daxiao Robotics (Chinese embodied AI startup) | H1 2026 Cumulative | Hundreds of Millions of USD · world-model

Daxiao Robotics disclosed cumulative fundraising of hundreds of millions of USD in the first half of 2026. Funds are primarily directed toward its "Kaiwu" world model and integrated hardware-software solutions, emphasizing human-centric ambient data collection to accelerate large-scale embodied training and on-device direct-drive control.Source: Kechuangtong Shares source (WeChat, CN)

III. Commercial Deployment

Galaxy General Robotics Galbot S1 Enters CATL Battery Cell Factory · humanoid

CATL (Chinese battery giant) has selected Galaxy General Robotics (Chinese humanoid startup), with the Galbot S1 heavy-load humanoid entering CATL's battery cell manufacturing lines for routine production operation in a new-energy manufacturing environment. Unlike a one-off demo, "routine operation" means the robot participates in production continuously on standard working hours — making Galaxy General Robotics one of the first humanoid players to enter a heavy-asset manufacturing facility, following leading automakers.Source: Interesting Engineering source

BMW Launches Figure 03 Logistics Sequencing Deployment (Spartanburg) · humanoid

BMW has activated the next-generation humanoid Figure 03 at its Spartanburg, South Carolina plant for complex logistics sequencing tasks: picking parts from unsorted bins, sorting and stacking them onto sequenced carts, and delivering them in order to assembly stations. This builds on an eleven-month Figure 02 pilot at the same plant's body shop — during which it assisted in producing over 30,000 X3s, handled over 90,000 parts, and logged approximately 1,250 work hours. Figure 03 adds a compliant outer shell, wireless charging, voice interaction, and an improved hand with tactile sensors and palm-mounted cameras. Both BMW and Figure AI position this as a milestone validating humanoids' transition "from lab to stable manufacturing workforce."Source: BMW Group source

General Motors Deploys 50 Collaborative Robots, Over 1,000 Positions Affected · industrial

General Motors has deployed approximately 50 collaborative robots at its factories, accompanied by workforce adjustments affecting over 1,000 workers. In contrast to the high media profile of humanoids, the at-scale replacement by mature collaborative robots on production lines more directly reflects the real-world labor impact of automation in manufacturing.Source: Storyboard18 source

Zhiyuan Robotics Spirit G2 Works 6 Consecutive Days on Consumer Electronics Line · humanoid

Multiple Spirit G2 units from Zhiyuan Robotics (Chinese humanoid startup) entered a third-party volume production factory (Longcheer Technology's Nanchang facility) for tablet quality inspection. Starting June 23, the robots worked for 6 consecutive days alongside human workers from 8 a.m. to 7 p.m. with a full livestream; the company claims a task success rate of 99.98%. ⚠️ vendor claim A public livestream is still a validating demonstration rather than a long-term mass production metric; the real test of whether this constitutes a deployable state is full-process compatibility, sustained stability, and low-cost replication — 6 days of data cannot be extrapolated.Source: Jiemian News source

IV. Industry Developments

Unitree R1 Retail Price Cut to ¥29,900 · humanoid

Unitree Robotics has lowered the official starting price of its R1 humanoid robot from ¥39,900 to ¥29,900, with stock available immediately and no waitlist. Following an earlier wave of "price disruptor" narratives at the start of the year, China's consumer- and research-grade humanoid market has pushed further below ¥30,000, shifting the competitive focus from specifications toward scale and cost.Source: 21 Caijing source

Amazon's Zoox Unveils Redesigned Robotaxi · autonomy

Amazon's Zoox has revealed a new-generation steeringwheel-free robotaxi with a boxy form factor and an emphasis on passenger comfort (featuring a panoramic roof), in preparation for upcoming commercial paid operations. This is a key product iteration ahead of its commercial expansion and will compete directly with purpose-built driverless platforms such as Tesla's Cybercab.Source: CNBC source

RoboScience Releases General Embodied Foundation Model Visics · world-model

RoboScience has released a general embodied foundation model, Visics, debuting a VLOA architecture and claiming cross-embodiment, cross-object, and cross-task execution capability at 1/200th the cost of conventional approaches. ⚠️ vendor claim The cross-scenario generalization and cost figures are self-reported from a launch event and lack independent reproduction.Source: Pandaily source

Zhiyuan Robotics Announces "Yuansheng" Ecosystem Plan and Dual World Model Track · world-model

Zhiyuan Robotics has announced dedicated investment of several billion yuan over the next five years to support a developer ecosystem for embodied AI deployment (the "Yuansheng" initiative), and has released five open-source projects. The company splits its world model into two tracks: a "World Action Model" (how actions change the world) and a "World Simulator" (an interactive, inferential, trainable environment). Co-founder Peng Zhihui stated at MWC Shanghai that production volume will exceed 15,000 units, with a new milestone to be disclosed once unit production surpasses 10,000 ⚠️ forward-looking statement.Source: Embodied Intelligence Lab source (WeChat, CN)

Hyundai Workers Plan Strike to Protest Humanoid Robots on the Factory Floor · humanoid

Following Hyundai's full acquisition of Boston Dynamics and plans to deploy Atlas humanoids in its US factories, some Hyundai workers are planning a strike to protest humanoid robots entering their workplace. This marks the first time organized labor has pushed back against humanoid robots in manufacturing, bringing labor relations into a debate that has until now centered on cost and efficiency.Source: Futu NiuNiu source

Hardware · Supply Chain

· Memory: Micron states that a single humanoid robot requires approximately 10× the memory capacity of an L2+ vehicle, and predicts that Optimus could become a larger memory customer than Tesla's cars — humanoid volume ramp-up is seen as a new growth driver for memory demand ⚠️ vendor/analyst claim (source: ifeng.com source).

· Precision Tubing: A company in Ningbo, China manufactures "vascular" tubing for robots, with a minimum wall thickness of 0.005 cm and annual capacity exceeding 1 billion units, entering upstream segments such as joint and liquid cooling circuits (source: Sohu source).

· Power Modules: Midea Welling (Chinese appliance and motor maker) has launched an integrated power module targeting humanoid joints, offering an integrated motor and drive solution aimed at reducing commercialization costs (source: Sina Finance source).

V. Weekly Observations

World's First Autonomous Driving Global Technical Regulation Approved; China's L3/L4 Mandatory National Standard in Final Approval Process · autonomy

At the 199th plenary session of the UN World Forum for Harmonization of Vehicle Regulations (WP.29), held in Geneva from June 22–26, delegates approved the Automated Driving Systems Global Technical Regulation (ADS GTR), jointly led by China, the EU, the UK, the US, Canada, and Japan — the world's first unified regulation of its kind. China served as vice-chair of the GRVA and co-chair of the ADS working group, with the Ministry of Industry and Information Technology (MIIT) organizing the China Automotive Technology and Research Center (CATARC) and the China Academy of Information and Communications Technology (CAICT) to lead the technical drafting. In parallel, MIIT has accelerated work on China's domestic mandatory national standards for automated driving systems; drafting is complete and the standards are now in the official approval process, with more granular safety requirements for L3/L4 and defined safety floors for each level. This provides a unified compliance reference for the autonomous driving industry, and signals that liability determination and market access will shift from self-reported claims to codified standards.Source: Guandian.cn source · MIIT/CATARC announcement

US Moves on Multiple Fronts Against Chinese Robotics: GUARD Act, Commerce Dept. Review, Pentagon Entity List · humanoid

The US government has taken a series of actions against China's robotics sector this month. Bipartisan legislators have introduced the GUARD Act, which would require national security agencies to review humanoid and quadruped robots manufactured in China and other countries. Commerce Secretary Lutnick told business executives on June 22 that the Commerce Department is reviewing imports of robots receiving state subsidies, and that the review could result in executive action steering federal procurement toward domestic manufacturers and restricting certain Chinese robots from sensitive facilities. The Pentagon added Unitree Robotics and other Chinese companies to its list of entities with military ties on June 8; the Department of Defense will be prohibited from renewing direct procurement contracts with those companies from June 30. ⚠️ many measures still at proposal/review stage Most actions beyond the entity list have not yet taken effect, but policy risk on the export and procurement sides has materially escalated.Source: House Select Committee on China · Washington Times source

China Accelerates Humanoid Quality and Component Standards: Maturity Evaluation Framework, Dexterous Hand National Standard · humanoid

Hangzhou Binjiang district has released China's first humanoid robot industry quality management system maturity evaluation standard, drawing on automotive industry experience and covering the full lifecycle from R&D to service. Separately, MIIT's "Humanoid Robotics and Embodied Intelligence Standards System (2026 Edition)" incorporates dexterous hands into the body and component standards, requiring a five-finger structure with at least three degrees of freedom per finger, providing top-level design guidance for industry standardization.Source: Hangzhou Binjiang Market Supervision source (WeChat, CN)

Morgan Stanley Doubles China Humanoid Robot 2026 Shipment Forecast to 50,000 Units · humanoid

Morgan Stanley has raised its forecast for China's 2026 humanoid robot shipments from 28,000 units to 50,000 (up from just 14,000 at the start of the year), citing a faster-than-expected shift from demonstration to commercial deployment. The bank projects annual shipments of 446,000 units by 2030, with market size growing from approximately $2 billion this year to $15 billion, implying a 2025–2030 CAGR of approximately 106%. Catalysts cited include a roughly ¥6.8 billion order from State Grid Corporation of China, policy support, and supply chain capacity expansion. ⚠️ analyst forecast This represents a single firm's projection, not an established fact.Source: Morgan Stanley Research source (WeChat, CN)

Industrial Robot Shipment Forecast: Annual Installations to Exceed 760,000 Units by 2030 · industrial

One industry forecast projects global annual industrial robot shipments will exceed 760,000 units by 2030, driven primarily by AI and automation. Separately, separate data indicate that global robotics funding in 2026 to date has already exceeded ¥60 billion, with capital and production capacity expanding in tandem.Source: Quality Magazine source · Xinji Tanke source (WeChat, CN)

This Week in Supply Chain: Six-Axis Force Sensors See Rapid China-Domestic Substitution; Dexterous Hands and Reducers Remain Bottlenecks · hardware

The most concentrated hardware signal this week is in six-axis force sensors — as humanoids move from "1–2 force-controlled joints" in industrial end-effectors to wrist, ankle, dexterous hand, and eventually whole-body force sensing, demand has structurally stepped up. Multiple white papers and research sources indicate: China's six-axis force sensor market exceeded ¥300 million in 2025, with shipments surpassing 10,000 units and year-over-year growth exceeding 300%; the China-domestic supply ratio has reached approximately 90%, with Changzhou Kunwei (Chinese force sensor maker) holding roughly a 53% share, strain-gauge designs as the mainstream, and MEMS (from companies such as Sensitron) as an emerging alternative. ⚠️ mostly vendor/institutional survey data Scope and methodology vary; absolute figures should be treated with caution. On the actuator side, competition in dexterous hands around lightweighting (PEEK materials) and "high-density motors" is intensifying. Reducers and joint modules are back in focus given reports that Tesla's Optimus V3 is targeting production start at Fremont in July–August with harmonic reducer and joint module sourcing from Taiwan suppliers. Overall, force sensing and transmission are the two segments of the humanoid supply chain under the most strain — and receiving the most capital and policy attention — this week.Source: Xinmaterial Online source (WeChat, CN) · Murong Daxia's Telescope source (WeChat, CN)

FutureX · Physical AI Daily — Issue 38 (06/25)

Shawn — Wed, 24 Jun 2026 14:50:59 +0000

Today's Highlights

· Agility Robotics plans to go public via SPAC merger with Churchill Capital XI at a $2.5 billion valuation, which would make it the world's first pure-play humanoid robot public company; Foxconn leads a $200 million PIPE, and its Digit robot already holds over $300 million in multi-year orders.

· Unitree (Chinese robotics company) R1 Air's starting price drops by ¥10,000 to ¥29,900 with immediate in-stock availability and no waitlist, further lowering the entry barrier for consumer- and education-grade humanoids.

· AGILINK (Chinese dexterous hand startup, spun out of Zhiyuan Robotics) closes a new round of nearly ¥1 billion, with a post-money valuation exceeding $1 billion — four rounds in five months, marking a clear capital shift toward "the hand."

· SEER Robotics (Chinese robot controller manufacturer), billed as "the first robot brain stock," lists on the Hong Kong Stock Exchange (06106.HK); its intelligent robot controllers have ranked #1 globally by sales for three consecutive years with roughly 25% market share.

· Morgan Stanley raises its China humanoid robot shipment forecast for 2026 to 50,000 units — nearly double its prior estimate — calling the industry's entry into "early commercialization."

I. Research Papers

SPACE: Bridging Cross-Robot Data with a Unified Action Representation · vla

Cross-embodiment transfer is a core obstacle to generalist policies — the same action maps to different motor commands across different robots, making behavior cloning data hard to reuse. SPACE instead uses "Cartesian state deltas" as a universal action representation across robots, with separate adaptation layers for embodiment differences, temporal scale, and execution commands, enabling demonstration data from heterogeneous robots to transfer to one another and converge toward a generalist policy.

Haeone Lee et al. (KAIST) · arXiv 2606.24049 source

World Value Models: Learning Value Functions via World Models Instead of VLMs · world-model

General-purpose value models are key to scaling large volumes of mixed-quality data for robot policy learning, but mainstream value models are built on VLM backbones pretrained on static or temporally sparse observations — lacking the temporal modeling required for value estimation. This paper instead builds value functions on world models, which are naturally suited to temporal reasoning and future planning, balancing grounding in historical context with prediction of future outcomes.

Zhihao Wang et al. (Tsinghua AIR) · arXiv 2606.24742 source

InSight: Enabling VLAs to Autonomously Acquire Missing Skills · vla

A VLA's capabilities are bounded by the skills present in its training data. InSight makes VLAs language-controllable at the level of "primitive actions" (e.g., "move gripper over bowl," "lift upward"), then uses a VLM-driven data flywheel to automatically identify which primitives are missing for a new task and autonomously collect demonstrations of those primitives — breaking through the skill boundaries of the training set.

Maggie Wang et al. (Stanford) · arXiv 2606.24884 source

Supervise What Survives: How to Use Synthetic Robot Video for VLA Training · vla

A common practice is to treat generated robot video as real data and back-project pseudo-actions from pixels. This paper argues this is a "mismatched abstraction": video retains only geometry (where the task happens), while real demonstrations contain control signals (exactly how motors move). It proposes an "asymmetric retention principle" — supervise only the geometric information that reliably survives the generation process when adapting VLAs, rather than forcing the model to learn low-level control signals that have already been lost.

Danze Chen et al. · arXiv 2606.24448 source

ReTVL: Treating Retries in Demonstrations as Supervision, Not Noise · manipulation

Human demonstrations often contain correction segments — missed grasps, misalignments, repeated attempts — that have typically been discarded as noise. ReTVL instead uses "retry events" as sparse supervisory signals to learn an error-sensitive value function, breaking from the common "monotonic progress" assumption to better characterize local execution errors and how task feasibility can be recovered.

Xinyao Qin et al. · arXiv 2606.24633 source

FT-WBC: Fault-Tolerant Whole-Body Control Under Actuator Failure · locomotion

Loco-manipulation robots face greater fall risk under actuator failure due to center-of-mass shifts and dynamic disturbances introduced by the arm, yet existing fault-tolerant methods mostly address locomotion without considering arm reachability. FT-WBC uses an upper/lower body decoupling strategy with a fault estimator to simultaneously maintain whole-body stability and arm reachability under actuator failure, extending fault tolerance from pure locomotion to loco-manipulation.

Yudong Zhong et al. · arXiv 2606.24466 source

FEARL: A Formally Verifiable Safety Module for Robot Foundation Models · vla

When foundation models are used for robot control, the same expressiveness that enables rich perception makes the model opaque and hard to formally verify. FEARL addresses this through modular decomposition: a large Controller handles high-dimensional perception and task reasoning, while a small Safety module receives only low-dimensional observations from dedicated safety sensors plus a bounded context embedding from the Controller, then produces the final action — keeping safety-critical requirements such as collision avoidance in an analyzable small module.

Davide Corsi et al. (UC Irvine) · arXiv 2606.23754 source

APR: Making Dexterous Hands Play Piano More Like Humans · manipulation

Reinforcement learning can achieve high accuracy piano playing for dexterous bimanual hands in simulation, but high-DoF hands optimizing for task reward alone often produce unnatural postures and joint hyperextension. APR uses adversarial pose regularization — drawing on a small amount of casual human playing data — to pull the policy's hand-pose distribution toward a human prior, avoiding expensive per-song demonstration alignment. The team also collected and open-sourced piano-playing hand motion data captured with consumer-grade Meta Quest 3, retargeted to the Shadow Hand.

Bin Qiu et al. · arXiv 2606.23848 source

Other papers today: NavWM (unified navigation world model for anticipatory planning, arXiv 2606.24101 source); DynaWM (world-model regularization for smooth wheeled-legged traversal of continuous stairs, arXiv 2606.24089 source); G³VLA (injecting camera calibration geometry into VLA visual tokens, arXiv 2606.24472 source); RoBoSR (object-level scene graphs as an intermediate structure for embodied reasoning, arXiv 2606.24338 source); Neuro-Symbolic Drive (supervising driving VLAs with reasoning traces from a rule-based planner, arXiv 2606.23938 source); ArtiTwinSplat (automatic reconstruction of interactive digital twins from RGB-D video, ETH, arXiv 2606.24628 source); TACTFUL (autonomous tactile-only exploration and object recognition, with Amazon participation, arXiv 2606.24712 source).

Open Source · Tools · Benchmarks

· TurboMPC: An end-to-end differentiable MPC solver running entirely on GPU (SQP + ADMM, co-designed JAX-CUDA implementation), supporting state/control inequality constraints and implicit integration. arXiv 2606.24039 source

· Tailor-Bench: A benchmark specifically testing visual world models' ability to simulate "long-tail and irregular physical interactions," directly addressing whether existing world models have truly internalized and generalized physical laws. arXiv 2606.24256 source

· REALM: A unified red-teaming evaluation benchmark for VLMs in the physical world, standardizing previously fragmented jailbreak/attack evaluations across datasets, metrics, and threat models. arXiv 2606.23892 source

· Amap DreamX-World 1.0 open-sourced: A 5B-parameter interactive world model generating at 16 FPS with streaming output, using geometry-retrieval memory to mitigate scene inconsistency during camera turns. source (WeChat, CN)

II. Funding & Deals

AGILINK ｜ New Round ｜ ~¥1 Billion ｜ Post-Money Valuation Exceeds $1 Billion · embodied

Mirae Asset, Jingming Capital, and a major internet conglomerate led or co-invested in the round, with existing investors including Hillhouse Ventures and Lanchi Ventures (Chinese VC firms) adding to their positions. AGILINK was spun out of Zhiyuan Robotics (Chinese humanoid robot company) in January 2026 to focus exclusively on dexterous hands; its founder Xiong Kun comes from HKUST's robotics institute and previously worked on Tencent Robotics X. The company says it has cumulatively delivered over 8,000 dexterous hands and over 10,000 grippers, with more than 1,000 units in routine industrial and logistics operation. Four rounds in five months, this deal is the most emblematic signal of capital shifting from complete robots toward "the hand." Source: 36Kr source

Agility Robotics ｜ SPAC Merger ｜ $2.5 Billion Valuation ｜ Over $620 Million Total Raise · humanoid

Merging with Churchill Capital Corp XI, controlled by Michael Klein; the post-merger ticker will be AGLT. The raise includes a $200 million PIPE at $10/share led by Foxconn (an existing shareholder), plus approximately $420 million from the trust account. The company disclosed that its Digit v5 already holds over $300 million in multi-year orders, has logged over 65,000 cumulative operating hours across 9 use cases, and has planned annual capacity of up to 10,000 units, with customers including Amazon and Schaeffler. If completed this year, it would become the world's first pure-play humanoid robot public company; the deal is also viewed as more substantive than most SPACs because all existing shareholders are rolling over their equity. Source: The Robot Report et al. source

Striding AI ｜ Angel Series ｜ ~$100 Million · humanoid

Charoen Pokphand Group, Huaqin Technology (Chinese electronics manufacturer), and 9AN Medical (Chinese listed company) participated. The company was founded in early 2026 by Yao Song, previously co-founder of Deep Lens (Chinese AI chip startup) and Orienspace (Chinese commercial launch company), with a focus on "physical intelligence" — aiming to advance humanoid robots into industrial and commercial settings via world-action models and reinforcement learning. Raising nearly $100 million in an angel round within six months of founding reflects the intensity with which industrial capital is backing early-stage embodied AI teams. Source: Sina Finance source

Tsinghua Post-2000 Tactile Team ｜ Angel Round ｜ Tens of Millions of RMB · hardware

Backed by Frees Fund and Pokeshell Robotics (Chinese robotics investor), the team develops ultra-thin visuo-tactile sensors targeting the tactile perception gap in dexterous hands and robotic arms — aligned with today's broader theme of capital flowing toward "the hand." Source: Hard Krypton source (WeChat, CN)

Suiqing ｜ Angel Round ｜ Tens of Millions of RMB · industrial

Focused on auxiliary mining robots for underground operations, founded by a team of Tongji University doctoral graduates targeting structurally underserved "blue ocean" environments such as underground mines. Source: Pedaily source

Embodied funding momentum continues: Kunlun Xing Robotics (Chinese humanoid startup) has completed three rounds in 90 days with existing investors continuing to add, and its valuation has already exceeded $1 billion (previously reported).

III. Commercialization & Deployment

Unitree R1 Air Drops to ¥29,900, Now In Stock · humanoid

The entry-level R1 Air's starting price falls from ¥39,900 to ¥29,900, with units available for immediate purchase and no pre-order queue required. The robot weighs approximately 27 kg, has 20 motion joints and monocular vision, and is aimed at education, training, and basic secondary development. Unitree says approximately 90% of core components — motors, reducers, and controllers — are developed and manufactured in China, which is the source of the cost reduction. Following earlier price cuts to the R1 series, the entry barrier for consumer- and education-grade humanoids has been lowered further, shifting from "pre-order waitlist" to "available off the shelf." Source: National Business Daily source

Galaxy General Robotics × CATL: Heavy-Load Humanoid with CATL Battery Deployed in Factory · embodied

Galaxy General Robotics (Chinese humanoid robot company) and Ningjiа Service — a subsidiary of CATL (Contemporary Amperex Technology, world's largest EV battery maker) — have signed a global strategic partnership; Galaxy's heavy-load industrial humanoid Galbot S1 (dual-arm maximum payload ~50 kg, industrial-grade spec) has achieved routine operation under complex conditions at a CATL battery factory. The deployment is on an actual production line, not a showroom, demonstrating that the "battery manufacturer + embodied robot" industrial collaboration is proving viable in heavy-load industrial settings. Source: Sohu source

Kawasaki Robotics Expands Partnership with Dexterity to Scale Physical AI in Warehouses · industrial

The two companies are expanding their collaboration to bring Dexterity's AI-powered picking and unloading capabilities to more warehousing and logistics settings, representing a substantive expansion in warehouse automation. Source: AI Insider et al. source

Zhiyuan Robotics (Chinese humanoid company) Takes Over an Entire 3C QC Line in 6-Day Live Stream · industrial ⚠️ Live demo framing

Zhiyuan used a continuous 6-day live stream to demonstrate its robots taking over a full 3C electronics quality-control line. President Peng Zhihui stated in parallel that "capital will no longer pay for robot demos alone — real-scene commercial value is required," and proposed leasing as the future mainstream model for robot deployment. The sustained live stream is intended to demonstrate stability, though the economics of continuous large-scale operation remain to be validated with production-scale data. Source: Xueqiu source

Boston Dynamics Announces ~$100 Million HQ Expansion, ~1,250 New Jobs · humanoid

Following Hyundai's full acquisition and plans for a production-ready Atlas to enter Hyundai's U.S. factories in 2028, Boston Dynamics has announced approximately $100 million in expansion investment at its Waltham, Massachusetts headquarters, adding approximately 1,250 jobs — another concrete milestone on Atlas's commercialization roadmap. Source: NBC Boston source

GM Factory Zero Brings in ~50 Robots as Over 1,000 Jobs Disappear · industrial

General Motors has deployed approximately 50 robots at its Factory Zero plant in Detroit to advance automation, with over 1,000 jobs eliminated in parallel — a concrete illustration of manufacturing's "robots replacing workers" dynamic. Source: WION source

Robotaxi International Expansion Continues: Pony.ai Enters Singapore, Caocao × May Mobility Expand into Europe · autonomy

Pony.ai (Chinese autonomous driving company) has partnered with a local mobility giant to launch in Singapore; Caocao Mobility (Chinese ride-hailing platform) and May Mobility have formed a strategic partnership to expand robotaxi services in Europe first; Waymo disclosed approximately 500,000 weekly rides. This continues the recent global expansion narrative (previously reported) — today's additions are new cities and partners, not a new paradigm. Source: Morningstar source / Sohu source

IV. Industry Developments

"First Robot Brain Stock" SEER Robotics Lists on Hong Kong Stock Exchange (06106.HK) · industrial

IPO priced at HK$101.6 per share, raising approximately HK$1.067 billion before the greenshoe, with a listing market cap of approximately HK$11.2 billion. The company is a platform-type robot controller manufacturer; according to Frost & Sullivan, its intelligent robot controllers have ranked #1 globally by sales for three consecutive years, with market share rising to approximately 25%, and controller gross margins exceeding 80%. Its founder is a Zhejiang University alumnus who left medicine to build robots. Together with Agility Robotics' U.S. listing announced the same day, "embodied robot IPOs" have moved from rumor to reality. Source: Industrysourcing / Sina source

Morgan Stanley Raises China Humanoid Shipment Forecast to 50,000 Units — Second Upgrade This Year · humanoid

Morgan Stanley has raised its 2026 China humanoid shipment forecast from 28,000 units to 50,000 units (up from 14,000 at the start of the year — the second upgrade this year), estimating China's humanoid market at approximately $2 billion in 2026 and $15 billion by 2030. The rationale is that the industry is transitioning from demonstrations to commercial use faster than expected, including a ~¥6.8 billion order from State Grid Corporation of China, supportive policy, and supply chain capacity expansion. This is a sell-side forecast, not realized shipment data. Source: Sina Finance source

Masayoshi Son Says SoftBank Has Started Mass-Producing Robots, Aims to Be "World #1" · humanoid ⚠️ Shareholder meeting statement

At SoftBank's annual shareholder meeting, Masayoshi Son said production robots have already started coming off lines at existing factories, that "there will soon be a formal announcement that will surprise everyone," and that SoftBank aims toward "physical ASI" and plans to complete its acquisition of ABB's robotics business within 2026. These are strategic statements and plans; the production scale and specific product form have not been independently verified. Source: 36Kr source

Alibaba Qwen Open-Sources "Language World Model" Qwen-AgentWorld · world-model

Available in two versions — 35B-A3B and 397B-A17B — using long-chain reasoning to simulate 7 categories of agent environments including MCP, search, terminal, SWE, Android, Web, and OS, along with a released evaluation benchmark AgentWorldBench; the 397B version scores 58.71, slightly above GPT-5.4 (58.25). To be clear, this is a world model for digital/software agents, not a physical embodied world model — but it reflects the "world model" paradigm extending toward closed-loop agent training. Source: ShuYuan AI source (WeChat, CN)

JD.com and Magic Atom Announce Strategic Partnership to Scale Embodied Robot Commercialization · humanoid

The two parties will jointly advance an industrialization model for embodied intelligence, targeting commercial deployment of humanoid robots in retail, logistics, and related settings. Source: Sina Finance source

Xiaomi YU7 GT Sets First Autonomous Driving Lap Record at Zhejiang Circuit: 1:49.434 · autonomy ⚠️ Manufacturer demo framing

Xiaomi says its YU7 GT autonomous driving system posted a lap time of 1 minute 49.434 seconds at Zhejiang International Circuit, claiming the first autonomous driving lap record at that track. This is a closed-circuit capability demonstration and is a different dimension from L4 deployment on open public roads. Source: Sina et al. source

Hardware · Supply Chain

· The supply chain behind Unitree's price cut: With approximately 90% of motors, reducers, and controllers developed and manufactured in China, Unitree's ability to compress full-system costs underpins the ¥29,900 price point.

· GlobalFoundries: Announced an AI chip solution for humanoid robots, with the upstream wafer foundry entering the embodied computing stack. source

· Kingfa Sci. & Tech. (Chinese specialty plastics manufacturer): Its modified engineering plastics are now supplying leading robot companies, with lightweight structural components entering mass-production supply chains. source

· Tendon-drive wire: 0.8 mm wire rope capable of bearing 50 kg loads — Chinese-made high-strength wire rope enabling the "grip strength" of dexterous hands and joints. source