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)
Top comments (0)