Today's Highlights
· Waymo launched paid robotaxi service in Denver, San Diego, and Tampa on the same day, growing its nationwide fleet to over 4,000 vehicles
· Mech-Mind (Chinese robotics vision company) listed on the Hong Kong Stock Exchange; 9 cornerstone investors subscribed $186 million, market cap exceeds HK$12 billion
· UBTech drew its first firm line, aiming for positive quarterly EBITDA in Q4 this year
· A "robot kindergarten" involving Turing Award winner Richard Sutton opened in Shijingshan, Beijing
· N₀-Foundation released 30,000 hours of visuo-tactile data, open-sourcing 5,000 hours of it
· Lucida turns real rooms into editable simulation assets, lifting scene F-Score from 0.794 to 0.924
Paper Progress
N₀-Foundation: 30,000 hours of visuo-tactile data, one-fifth open-sourced · perception
In embodied manipulation, tactile data has always been the hardest and most expensive category to collect at scale, and this work pushes the scale directly to over 30,000 hours of synchronized visual-tactile demonstrations across 6 embodiments, covering 450 tasks, with 5,000 hours released publicly as OpenNeoData. The team also released a visuo-tactile sensor, a tactile-capture version of the UMI interface built specifically for tactile collection, a cross-sensor transferable representation model called NeoForce, and two evaluation suites — NeoReal for real-robot testing and NeoSim for simulation testing. Their experiments aim to show that what policies actually benefit from is the underlying physical contact state itself, not signal appearances specific to any one sensor.
NeoteAI Team, Fudan TEAI Team · arXiv 2608.29601 source
Lucida: breaking cluttered rooms into individually manipulable simulation assets · perception
Real-to-sim pipelines typically demand precise instance geometry and unoccluded viewpoints from the outset — conditions real-world footage simply can't provide. Lucida keeps the overall "parse–generate–place" order but pushes the precision requirement to the end of the pipeline: it first parses video into a scene graph backed by multi-view evidence, then generates a complete asset for each instance separately, and finally hands off to GizmoAct — a VLM policy that treats object placement as multi-turn GUI operation, dragging object gizmos itself and judging when alignment is achieved. Tested on R2S-Scene, mAP came in 69% higher than Boxer, and scene F-Score rose from SAM3D's 0.794 to 0.924. This work has reached a popularity score of 53 on the HF community.
Minghan Qin et al. · arXiv 2608.30821 source
LightNav-0: no dedicated prediction head for navigation tasks · autonomy
Instruction following, open-vocabulary object search, and visual tracking have traditionally each required their own separate module. LightNav-0 instead folds all three into a single compact model through a unified token interface: a dual-channel pointing representation conveys intent in a way that is independent of task, scene, or embodiment, while a residual vector-quantized action tokenizer translates that generic intent back into a trajectory a specific embodiment can execute. Trained on a corpus spanning over 2,000 scenes and 4,000+ hours of data, the model achieves SOTA monocular success rates across 10 public navigation simulation settings, and demonstrates zero-shot cross-embodiment transfer on real robots. This work has a popularity score of 24 on the HF community.
Shaoan Wang et al. · arXiv 2608.30935 source
Hydra: planning entirely within a discrete latent space, no decoding back to pixels · world-model
World models are good at imagining the future, but using them for real-time control runs into a fundamental mismatch: the planner and the generative model don't operate on the same manifold, so every candidate has to be decoded into a high-dimensional image before it can be scored — prohibitively costly on real hardware. Hydra instead compresses visual state, physical pose, and control actions into a single shared latent space, then splits it with a per-modality vector-quantized bottleneck into a discrete vocabulary, so candidates can be ranked in place by a Kinematic-Perceptual Cost without ever generating pixels; the selected intent is then unrolled into a continuous trajectory by conditional flow matching. Tested on two real robot platforms, goal-directed planning outperformed existing world models, and closed-loop execution matched or exceeded mainstream reactive base policies.
Mohammad Nazeri et al. · arXiv 2608.28995 source
Blind Dexterity: a blindfolded Unitree G1 performs whole-body manipulation · locomotion
No cameras, no torque sensors, no tactile skin — the only available signal is joint encoders. With just that, the policy enables Unitree G1 to walk while being pushed (without even relying on IMU feedback), stop a soccer ball with its foot, lift a suitcase by feeling for the handle, and step onto an arbitrarily placed skateboard. The authors' explanation is that the way joint readings shift during compliant contact itself constitutes a whole-body tactile channel — the robot actively induces contact to probe its surroundings, and object pose can be decoded from just a short window of proprioceptive history. This effectively inverts the default assumption that expensive sensors have to come first.
Aditya Bhatt et al. (Jan Peters's group) · arXiv 2608.29487 source
A humanoid robot swings from a bar, succeeding 14 of 15 times · locomotion
Compared to flat ground, sparse 3D structures impose an entirely different perception challenge — the bar is thin and suspended in air, giving lidar very few returns. This system feeds raw scans from a head-mounted solid-state lidar directly to the policy, uses an attention encoder with recurrent memory to extract geometric information, and stitches together the jump-up, swing, and jump-down phases of expert motion via staged teacher-student training. To transfer to real hardware, the team explicitly modeled lidar noise, battery voltage sag, and actuator thermal limits during training, and swapped the end effector for a passive hook. Across three bar-spacing configurations, the full sequence was attempted 15 times and succeeded 14 times, with swing speeds reaching 0.5 m/s; using the same perception backbone to train a separate policy, the robot could even swing through an overhead bar with a cross-section of just 2 cm.
Efe Ongan et al. (ETH Zürich) · arXiv 2608.29769 source
Motus2: one set of weights serving as policy, simulator, and judge · world-model
Common practice has been to bolt an action head onto a world simulator, but prediction and decision-making never actually formed a true closed loop. Motus2 exposes three interfaces from a single shared-weight model: a policy proposes candidate action chunks, a simulator predicts the visual consequences of those actions, and a value model scores the predictions — together closing a self-improving decision-and-learning loop. Failed and suboptimal interactions, previously discarded, are now fed back into dynamics modeling and value learning. The data side was also built up progressively: starting from large-scale monocular first-person video, then moving to synchronized binocular data, then robot-domain adaptation, and finally landing on a biomimetic platform equipped with binocular vision, dual arms, dual dexterous hands, and tactile sensing.
Hongzhe Bi et al. · arXiv 2608.30237 source
RoboPhys-3D: separating "looks right" from "is right" in world models · benchmark
Video world models are increasingly used as data engines and simulators, yet no one had systematically verified whether their generated rollouts actually preserve 3D scene state, or whether they translate into genuinely executable actions. RoboPhys-3D is built on top of RoboTwin 2.0, spanning 50 manipulation tasks, 5,000 trajectories, and 25,000 multi-view ground-truth videos; both generated and ground-truth videos are passed through the same 3D reconstruction pipeline, so reconstruction error and generation error can be disentangled. The evaluation suite comprises 50 metrics across four tiers. Among four representative models, Cosmos 3 achieved the highest RoboPhyscore at 0.6330, or 92.7% of ground-truth level; but the failures exposed by state-level and execution-level metrics went undetected by both perceptual metrics and VLM judges.
Tianyi Wang et al. · arXiv 2608.28718 source
A single poster shifts autonomous-vehicle camera–lidar calibration by 33.9° · autonomy
Online calibration was designed to let vehicles self-correct for vibration- and temperature-induced sensor drift without a shop visit — but this paper turns it into a new attack surface. ACA uses a single adversarial poster to first fool the miscalibration detector into triggering the calibration process, then steer the calibration estimator toward incorrect extrinsics, with the poster's geometry and texture jointly optimized for both effects. On KITTI and nuScenes, the attack produces up to 33.9° of average rotational error, collapsing downstream object detection; in CARLA simulation, once a poisoned calibration is accepted by the system, it leads to collisions; and a printed version of the poster attached to a real Husky robot reproduced the same errors. Incorrect calibration propagates through the fusion chain all the way into planning and control.
Liangkai Liu, Qingzhao Zhang, Kang G. Shin (University of Michigan) · arXiv 2608.28778 source
Other papers today: Matrix-Game 3.5 uses a patch memory mechanism to let real-time interactive world models sustain minute-scale long-horizon consistency (arXiv 2608.29910 source); NavMCP separates the roles of the VLM reasoning agent and the navigation foundation model, outperforming step-by-step interfaces by 14.9 points on HM-EQA (arXiv 2608.30396 source); DriftingVLA generates an entire action chunk in a single forward pass, eliminating the online latency of multi-step flow refinement (arXiv 2608.29749 source); AnyWorld can expand a single human interaction into multi-embodiment, multi-view, multi-scene robot trajectories (arXiv 2608.29242 source); DREAM generates fine-tuning data on-site via real-to-sim at deployment time, eliminating the need for human demonstrations (arXiv 2608.29078 source); SpectraTac built a compact camera-free optical tactile sensor using RGB active illumination with distributed color sampling (arXiv 2608.30368 source); one study found self-play driving policies exhibit reward hacking at traffic lights and have no inherent incentive to stop at stop signs (arXiv 2608.30819 source); drifting behavior turns out to emerge organically from lap-time optimization pressure, with no dedicated drift reference built into the reward design (arXiv 2608.28723 source).
Open Source · Tools · Benchmarks
· GHOST: an open-source VR teleoperation system letting a single operator control two mobile manipulators simultaneously, stitching together a third-person 3D workspace entirely from onboard RGB-D point cloud alignment (arXiv 2608.29080 source)
· Agri-Sim: a Unity + ROS2 greenhouse agricultural robotics simulation platform enabling closed-loop, repeatable evaluation of navigation, motion planning, and manipulation tasks (arXiv 2608.29100 source)
· DARP: an RGB-D-IR calibration dataset for dual-arm eye-in-hand setups, with the two arms photographing each other from opposite sides of a table, specifically designed to address self-occlusion under single-view capture (arXiv 2608.31002 source)
Funding and Deals
Mech-Mind (Chinese robotics vision company) | Hong Kong Stock Exchange main board listing | Raised HK$2.35 billion | Market cap over HK$12 billion · embodied
Following price-setting completed last week, Mech-Mind (09615.HK) listed formally on September 1, pricing at the top of its range at HK$101.7 and issuing 23.14059 million shares; if the over-allotment option is fully exercised, total proceeds could reach HK$2.7 billion. Nine cornerstone investors together subscribed $186 million, accounting for 62% of the offering: Baillie Gifford subscribed $60 million, Taikang Life $40 million, Jane Street, Invus, Ghisallo, and Ruihua each subscribed $15 million, NGS Super Fund and E Fund each subscribed $10 million, and BYD's Golden Link subscribed $6 million. The company makes standardized "eye-brain-hand" components; according to Frost & Sullivan figures, in 2025 it held a global shipment share of over 27% in the AI+3D-vision-guided general-purpose robotic component market — more than its next four competitors combined. Revenue grew from RMB 181 million in 2023 to RMB 389 million in 2025, with gross margin over the same period rising from 39.1% to 64.6%, and adjusted net loss narrowing to RMB 109 million; overseas revenue share rose to 50.3%, exceeding domestic (China) revenue for the first time in 2025. Although the public offering was oversubscribed roughly 3,835 times, the stock opened flat on its first day and at one point fell more than 4% intraday. Founder Shao Tianlan told LeiPhone that the annual shipment volume of general-purpose robots actually deployed in productivity settings is still only in the low hundreds of thousands, "several thousand times short of true mass adoption."Source: LeiPhone source
Reframe Systems | New round | $40 million · industrial
The round was led by Energy Impact Partners, with participation from Counterpart Ventures, E12 Ventures, Global Brain, Thin Line Capital, Up Partners, LACI Impact Fund, and others, alongside continued investment from existing backers Eclipse and RA Capital. All three co-founders — Felipe Polido, Aaron Small, and Vikas Enti — came from Amazon Robotics; the company was founded in 2022 and focuses on small, demand-proximate automated home-building factories. The company says its delivery speed is three times that of traditional construction, at 35% lower cost; it has delivered 10 homes to date, plans to deliver 114 more over the next year, and has set the capacity ceiling of its next factory, soon to come online in Massachusetts, at 500 multi-family units.⚠️ Company claimSource: The Robot Report source
Tiangong Robotics (Beijing Tianxiaxian Zhichuang) | New round | Hundreds of millions of RMB · industrial
The round was led by Chuxin Fund, with participation from Matrix Partners China and Hefei state-owned capital. Founder Xiao Jun previously served as JD.com Group vice president and president of its X division, where he led the company's unmanned warehouse, driverless vehicle, drone, and specialty robotics businesses. Founded in 2022, the company focuses on logistics sorting, has grown to a team of over 200, with roughly 85% in R&D roles; its flagship products are the "Wooden Horse" sorting robot and an expansion module that increases a single unit's sorting slot count to 4-5 times the original, and its business now extends to Japan, South Korea, the U.S., and Europe. The funds will go toward expanding sorting-robot production, developing and commercializing sorting-oriented humanoid products, and expanding into overseas markets.Source: ZhiDongXi source
Physical Superintelligence (PSI) | Seed round | $58 million · adjacent
The round was led by Breakthrough Energy Ventures, with participation from Dragon Global, Robot Ventures, SV Angel, Valkyrie, and others, plus individual investors from OpenAI, NVIDIA, Oracle, and Hugging Face. The company was formally founded on September 1 in Cambridge, Massachusetts, positioning itself as an AI-native physics lab staffed by "virtual physicists," with its core platform Emmy breaking research questions into a verifiable hypothesis tree and testing them in parallel. Its first chosen commercial application is AI data centers, using physics-native reasoning and simulation to solve multi-physics design problems spanning power, cooling, networking, and compute. CEO Matt Pines said PSI's mission is to "industrialize the discovery of new physics."Source: Unite.AI source
Youlichi | Strategic investment | Tens of millions of RMB · embodied
The investor is Keli Sensing, a publicly listed force-sensor maker; the investment is in the tens of millions of RMB. This is another stake in Keli Sensing's ongoing strategy of taking positions along the robotic perception supply chain.⚠️ Single-party claimSource: Robot Lecture Hall source
Locus Robotics | Series G | Nearing close · industrial
The warehouse-robotics and related-software company's Series G round is nearing completion; specific amount and investors have not yet been disclosed.Source: Axios source
Commercial Deployment
Waymo launches in three cities in a day; Zoox pushes testing into Houston · autonomy
Denver, San Diego, and Tampa all opened to public paid robotaxi service on the same day, September 1, with Colorado marking Waymo's first commercial operation in that state. A company spokesperson said each of the three cities will start with a few dozen vehicles, "growing to the hundreds over time," with rides booked through the Waymo app. The nationwide fleet now exceeds 4,000 vehicles across 14 cities, delivering over 500,000 paid rides per week, with the company aiming to push that past one million rides per week by the end of 2026; the new Ojai model is built on a Zeekr chassis, assembled at the Arizona factory, and uses sixth-generation autonomous driving hardware. Zoox, meanwhile, said it will begin safety-driver testing in Houston and San Diego, bringing its footprint to 12 markets, though its only paid service currently still operates solely in Las Vegas, with its San Francisco service still limited to free rides for select riders. The near-simultaneous moves from both companies land just ahead of Tesla's Thursday unveiling of the Cybercab and Robotaxi. Goldman Sachs research projects the U.S. robotaxi market will reach $19 billion by 2030. Labor groups worry about job displacement, while vehicle-safety advocates are calling for companies to disclose mileage and collision data using consistent, comparable standards.Source: CNBC source
Zhipingfang's AlphaBot mixes drinks at a Hong Kong Lan Kwai Fong bar · embodied
Starting work formally on August 31, the AlphaBot 2, equipped with the brain-inspired embodied foundation model NeuroVLA, took its place behind a real bar counter to mix cocktails for customers. More notable than the robot itself is the operational pipeline behind it: hardware customs clearance, regulatory compliance, obtaining a local Hong Kong beverage sales license, and building a paid service system — only once that full chain is working can this be called routine operation rather than a one-off demo. Lan Kwai Fong, home to over a hundred bars and restaurants, provides a ready-made stress test through its fluctuating foot traffic, noise, and lighting. Manufacturer-reported data shows motion jitter reduced by more than 75%, reflex response after a collision completing within 20 milliseconds, and power consumption of about 0.4 watts. Zhipingfang's earlier modular service space "Smart Cube" is already operating in over a dozen provinces and cities, with robot staff there making coffee, ice cream, and matcha.⚠️ Company claimSource: PEdaily source
GreyOrange's in-store system now runs in 3,800 retail locations · industrial
GreyOrange technology is now running in over 3,800 retail stores worldwide, with major retailers including the H&M Group using its gStore for inventory and shelf management, having tracked roughly 200 million items cumulatively. The system is sensor-agnostic, working with either overhead or handheld RFID readers, and can pinpoint an item's location to within 3 to 5 feet. The company disclosed that 130,000 software and hardware agents are currently running across its stores and warehouses, with 100,000 of those added in the past two years.Source: GlobeNewswire source
Tesla's Robotaxi fleet grows to nearly 200 vehicles in three weeks · autonomy
Data compiled by the crowdsourced platform Robotaxi Tracker shows the number of Tesla vehicles operating without a safety monitor across Austin, Dallas, and Houston has grown to nearly 200, roughly a 7x increase in three weeks; regulatory registrations for the two-seat Cybercab in Texas also jumped from 7 to 45 within a few days. The fleet had long held steady at around 20-odd vehicles, leading critics to question whether Tesla's actual deployment capability matched its stated goals. Autonomy lead Ashok Elluswamy said on the Q2 earnings call that the Robotaxi program has completed over 380,000 miles of unsupervised driving with no notable incidents. Taken together, these two developments suggest Thursday's event is likely more than just a simple new-car launch.⚠️ Aggregated dataSource: Teslarati source
Singapore MRT station trials humanoid guide robot, it mishears a place name · humanoid
Singapore's Land Transport Authority (LTA) placed a 1.27-meter tall Unitree humanoid robot named Olly at Little India MRT station for a two-week trial, running weekday mornings from 10 a.m. to noon, from August 31 to September 11. Its performance on day one was mixed: it handled simple questions like first/last train times fine, but when asked how to get to Tekka Centre, it needed multiple prompts before giving verbal directions, couldn't lead the person to the nearest exit, and at one point misheard the name as "Teacher's Centre," pointing the wrong way. When asked how to reach KK Women's and Children's Hospital, though, it correctly identified Exit F and proactively led the way all the way to the gantry. Steve Walker, a 69-year-old retiree with a robotics background who came specifically to try it out, said the robot's balance and gait were stable, but that it struggled to distinguish voices from background noise in the noisy station hall, and suggested adding a directional microphone. LTA said it chose off-peak hours for the trial and currently has no deployment plans.Source: CNA source
T-Robotics wins additional Ford ESS line automation order · industrial
The contract is worth approximately $2.63 million (roughly KRW 3.6 billion), supplying Ford's energy storage battery production facility in the U.S., covering automation equipment, installation and commissioning, and control-system development integration. Combined with the roughly KRW 15 billion AMR contract from this past May, cumulative orders tied to Ford's ESS line had reached about KRW 19 billion as of September 1. Ford's production line is being built in phases, leaving room for further orders.Source: Asia Economy source
Industry Developments
UBTech's revenue doubles in H1, gives first quarterly break-even timeline · humanoid
In the six months ended June 30, 2026, the company's revenue reached RMB 1.269 billion, up 104.2% year-on-year; gross profit reached RMB 567 million, up 160.9%, lifting gross margin by 9.7 percentage points to 44.7%; the period's net loss narrowed 23.0% to RMB 339 million. Following last week's disclosure that revenue from its full-size models had grown 1445%, the company said for the first time on its August 30 earnings call that it now aims to bring forward its target of quarterly break-even — originally set for Q4 2027 — to positive Q4 EBITA this year. CEO Zhou Jian noted that in both 2024 and 2025, deliveries in the second half outpaced the first half; while the Guangxi factory won't come online until mid-to-late September due to natural conditions, the company remains confident in its full-year revenue target of RMB 3.5–4 billion or more. The revenue mix has shifted sharply: full-size embodied-intelligence humanoid robots brought in RMB 590 million, jumping from 6.1% to 46.5% of total revenue year-on-year, with H1 sales of 921 units; revenue from other smart robot products, including educational robots, was RMB 246 million, down 18.8% year-on-year. Cash also fell sharply, from RMB 4.888 billion at the end of 2025 to RMB 2.326 billion, mainly due to a RMB 1.665 billion cash payment for a 43.01% stake in FengLong Co. First-half credit impairment losses reached RMB 91.07 million, versus just RMB 1.3 million in the same period last year, which the company attributed to a change in expected credit loss methodology from an external-rating approach to a migration-rate approach.Source: Nanfang Daily Wancaishe source
Skild AI releases S1, replacing spoken instructions with a video · world-model
Rather than a text prompt, this time it's a video. S1 positions itself as an "in-context learner": given a video of a human performing a task, without any fine-tuning or weight updates, it can directly produce robot actions, for tasks up to 10 minutes long — including repotting a plant, flipping pancakes, pour-over coffee, and kit assembly, none of which appeared in its pretraining data, per company demos. The company's blog reports 96% success on in-distribution tasks and 66% on out-of-distribution long-horizon tasks, versus just 9% for a comparably-sized language-prompted baseline; under noticeable perturbations, the language-model approach degraded three times as much as S1. This directly targets the dominant current paradigm, where every new task requires fine-tuning a dedicated policy on tens to hundreds of hours of teleoperation data. Skild was founded in 2023 and has raised nearly $1.7 billion to date, having just closed a $1.4 billion Series C last week (previously reported).⚠️ Company claimSource: The Robot Report source
"Robot kindergarten" opens in Shijingshan, Beijing, with Sutton's team co-building · embodied
It opened formally on September 1, built jointly by Tashan Technology and a team led by Turing Award winner and reinforcement-learning pioneer Richard Sutton, focused on tactile perception and continual learning. It's explicitly designed to bypass the current dominant approach — rather than relying on human demonstrations, robots repeatedly try things on their own, and when they hit a wall, that failure gets written into the learning data before trying a different way. Sutton described the goal as "providing a safe environment for a robot to learn about its own body and how to interact with the physical world through experience," with "failure being part of the process." One spider-shaped small robot in the facility learned to move forward in about 40 minutes with no prior training — though this remains a very narrowly defined task. For now, this approach looks more like a complement to teleoperation and imitation learning than a replacement for it.Source: Global Times source
Goldman Sachs raises 2030 humanoid shipment forecast 3.5x · humanoid
In a newly released "Physical AI" research report, Goldman Sachs now projects global humanoid robot shipments will reach approximately 890,000 units by 2030 and about 6.5 million by 2035, corresponding to a market size of roughly $138 billion — up from previous forecasts of 256,000 and 1.4 million units, respectively. This year's baseline shipment forecast was also raised from 51,000 to 75,000 units. The report identifies logistics and warehousing as the leading early-adoption use case: Amazon has already deployed over one million robots across more than 300 sites, and Walmart's freight automation now covers 3,100 stores nationwide; Goldman estimates automation could save Amazon roughly $72 billion in service costs by 2030. On chips, the report estimates chip content per humanoid robot at over $3,000-$6,000. Another easily overlooked finding concerns factory control systems: as humanoid robots proliferate, PLCs are shifting toward software-defined vPLCs, a market growing 20-30% annually, posing a challenge to established vendors like Siemens and Rockwell.⚠️ EstimateSource: Guandian.cn, citing Goldman Sachs research report source
Hugging Face's rubber-duck robot sells out fast, chip sourced from Rockchip · adjacent
The Microduck, launched by Hugging Face's French subsidiary Pollen Robotics, has sold over 10,000 units since going on sale last Thursday; at $399 each, that's already over $4 million in revenue, and delivery of new orders has slipped past the originally promised Christmas 2026 date. The 800-gram duck-shaped robot doubles as both toy and development platform, running on the RK3566 chip from Shanghai-listed Rockchip (Chinese chipmaker), which uses licensed ARM technology. Omdia chief analyst Lian Jye Su said Rockchip is a key supplier in edge AI, with its chips commonly used in machine-vision applications like object detection and image recognition, noting it's "very widely deployed, but with limited compute resources — not purpose-built for sophisticated edge AI devices." Rockchip's H1 revenue reached RMB 2.88 billion, up 40% year-on-year, with non-GAAP net profit growth exceeding 60%. Several other players have already entered this price segment: Zeroth's child-sized humanoid robot is priced at RMB 8,888 and has taken 247 pre-orders on JD.com.Source: CNBC source
500+ China A-share companies mention "embodied intelligence" in half-year reports, single digits disclose revenue · adjacent
China Fund News combed through 2026 half-year reports and found over 500 A-share companies mentioning "embodied intelligence" in their filings, but only 23 discussed it in any depth (10+ mentions), and the number actually disclosing corresponding revenue figures was in the single digits. The companies that could actually back it up with numbers share one trait: their core business already had a natural interface with embodied intelligence. Sensor and control maker Sinexcel renamed its product segment from "Smart Devices" to "Embodied Intelligence" — its frameless torque motors are now being supplied in volume to several leading manufacturers, with segment revenue of RMB 60 million, up 114.7% year-on-year; Dobot's (Chinese robotics maker) embodied-intelligence business revenue surpassed RMB 45 million, up over 20x year-on-year, accounting for 14.3% of total revenue. On the other end are companies with struggling core businesses: Hoson Intelligent posted an H1 net loss of RMB 425 million yet mentioned embodied intelligence 24 times in its half-year report; Shangpin Home Collection saw revenue decline 24.2% and posted a net loss of RMB 233 million, while also touting a high-profile push in this direction. Mention frequency didn't correlate with actual performance either way — Kejie Intelligent's net profit grew 251.6% while mentioning the term 35 times, and Kepu Cloud mentioned it 63 times. The sector saw 322 financing deals in H1 totaling roughly RMB 93.5 billion, already exceeding the full-year 2025 total.Source: China Fund News source
Unitree's market cap down about RMB 218.8 billion from its listing-day high · humanoid
Shares hit an intraday low of RMB 555.8 on August 31, a new post-listing low, closing the morning session at RMB 559.11, down 4.43%, for a market cap of RMB 226.14 billion — a 49.17% pullback from the RMB 1,100 opening price on its first day of trading. The stock recovered 1.1% on September 1, with the company also denying, on the same day, earlier rumors concerning expense-reimbursement approvals.Source: China.com Finance source
Galbot's ET1 opens pre-orders September 3 · humanoid
This is the company's first small bipedal humanoid robot, standing 1,230 mm tall and weighing 30 kg, with a flexible-material shell and a motor solution supplied by Unitree. It runs Galbot's general-purpose "cerebellum" foundation model, AstraBrain-WBC, trained on 100,000 hours of human motion data using a mix of high-precision motion capture and video retargeting.⚠️ Company claimSource: Gasgoo source
Hardware · Supply Chain
· Folee New Materials → Linker Bionics: cumulative deliveries of tactile sensors have now exceeded 30,000 units, primarily fitted to the industrial dexterous hand Linker Hand O6; under the two companies' February agreement for a total procurement of 100,000 units, over 30% of the order has now been fulfilled source
· QJ Robotics whole-hand tactile 2.0: extends tactile sensing from the fingertips across the finger pads and palm, with fingertip marker density reaching 125 points/cm², combined-force precision of 0.03N, and a refresh rate of 120Hz — 4x the previous generation; the flexible e-skin offers force resolution of 0.01N with a range up to 200N; a hub board in the palm can connect 5 sensor channels and packs up to 6 TOPS of on-device compute, cutting the host's perception compute load by up to 90% source
· Robot Components Industry Development Alliance: launched by JD Industrials together with Schaeffler and others at the 2026 World Robot Conference, aiming to use real industrial use cases to drive next-generation component performance, lifespan, and safety improvements, and to bring costs down through standardization and scale. Schaeffler Greater China e-drive division president Chen Xiangbin said the alliance "bridges the gap between components and actual robot deployment" source
Top comments (0)