Today's Highlights
· Waymo unveils its first self-developed robotaxi chip — built on TSMC 5nm, delivering over 1000 TOPS, already deployed
· Nevada approves 8,000 vehicles for commercial Robotaxi operation in one go; Tesla alone capped at 5,000
· Hyundai Motor Group divests its entire 47.5% stake in RAI (Robotics and AI Institute, spun out of Boston Dynamics), buyer is SoftBank
· 49 Chinese central state-owned enterprises jointly attend WRC with procurement lists for the first time; State Grid's RMB 6.8 billion embodied-AI equipment procurement enters supplier matching
· China's State Administration for Market Regulation folds embodied intelligent robots into phase two of its quality supply-chain strengthening initiative, rolling out 166 tasks
· In H1, 60% of humanoid robot shipments went to performance and research use; under 20% entered core production settings
· Tactile prediction lifts real-robot success rate on contact-intensive manipulation from 31.1% to 72.2%
Research Progress
HiTac-WAM: "Rehearsing" Tactile Sensation Before Acting Doubles Success on Contact-Intensive Tasks · manipulation
World models have typically represented predicted future tactile signals as a single image or latent vector; HiTac-WAM instead uses hierarchical prediction, decomposing the predicted tactile future of each candidate action chunk into three tiers — contact state, 3D deformation field, and slip risk — where each downstream tier only receives stop-gradient signals from the tier above, and replanning is triggered during execution whenever predictions persistently diverge from real readings. Across three tasks — chip grasping, whiteboard erasing, and USB insertion — using hierarchical prediction to select actions raised the average real-robot success rate from 31.1% to 61.1%, with the full system reaching 72.2%; compared to a baseline that predicts deformation only, 3D displacement error dropped by 17.6%.
Chao Xue et al. · arXiv 2608.19574 source
What Actually Matters in Latent Action Models: A First Unified Benchmark Across 41 Design Choices · vla
Latent Action Models (LAMs — which compress actions from unlabeled video into "proxy actions" that robots can learn from) have seen a year of divergent methods and incomparable experimental setups. This work unifies representative methods under a single autoencoding framework and systematically sweeps 41 design choices across three axes — modeling paradigm, learning objective and regularization, and how latent actions are fed into downstream policies — while also testing whether four commonly used proxy metrics can predict downstream manipulation performance. Results across three standard benchmarks plus real-robot validation converge on one consistent finding: fine-tuning a VLM backbone with latent actions provides a stronger initialization for downstream policy learning.
Xizhou Bu et al. · arXiv 2608.19613 source
Swap the Robot, Lose the Skill: Letting a VLA Generate Its Own Demonstrations to Learn New Skills Without Forgetting Old Ones · vla
VLAs like π0.5 suffer large performance drops when transferred to a new robot embodiment, even with only minor hardware configuration differences; fine-tuning on expert data from the new embodiment, in turn, tends to erase the model's prior instruction-following and behavioral priors. This method has the zero-shot VLA run interactive trajectories online first, then mixes these "self-demonstration" episodes into the fine-tuning set — the resulting policy inherits old-task knowledge distilled from the zero-shot model, preserves general instruction-following, and learns new skills with higher sample efficiency. Validated on the real-robot ALOHA platform and the new RoboTwin simulation benchmark.
Prachi Garg et al. · arXiv 2608.19490 source
Teleoperation Still Can't Match Human Hands — the Missing Piece Is "Distributed Tactile Feedback" · perception
Teleoperation is currently the primary means of collecting demonstration data, but even on basic tasks, teleoperated performance remains notably worse than direct human manipulation. This work provides evidence using a two-DoF bilateral force-feedback teleoperation device paired with a 32-DoF fingertip tactile display: faithfully reproducing local deformation from the remote gripper onto the operator's fingertips not only speeds up tasks but reduces corrective actions and step counts, cutting the deviation between teleoperated trajectories and natural human motion by 29–79%. Further increasing the quantization resolution of tactile feedback also compresses the state-space distribution of teleoperated actions — a feature previously believed to correlate positively with autonomous policy training outcomes.
Rohan Kota et al. · arXiv 2608.19372 source
AdaPT: Learning Professional Tennis Serves and Rallies Directly From Broadcast Footage · locomotion
Humanoid robots playing tennis is not new, but the hard part is preserving task performance while achieving "professional-looking" style. AdaPT uses a hierarchical design — a planner that generates stylized kinematic trajectories (sourced directly from broadcast video) and a tracker that executes them; to address sim-to-real tracking decay, the tracker is trained to follow randomized execution speeds, and the planner gets a learned motion-speed adapter to suppress error accumulation. Real-robot experiments were validated on Unitree G1 and deployed to the 1.7-meter full-size humanoid Dobot Atom, achieving mocap-free outdoor serving.
Tao Huang et al. · arXiv 2608.20087 source
DECOWAM: Separating the Camera's Own Motion From Base and Arm Actions · world-model
Mobile manipulation requires jointly predicting how walking and grasping will change future visuals and control, but existing world-action models are mostly designed for fixed bases and don't distinguish camera ego-motion from base and arm actions. DECOWAM freezes the FastWAM backbone and trains only a residual adapter, using an adversarial approach to disentangle base and arm latent variables, and conditions video prediction on base velocity; it also releases a companion real-robot dataset, ARMDOG, with synchronized video, whole-body state-action, and language annotations. Under a fixed playback protocol, action MSE dropped by 21.7%, with only 25.95 million trainable parameters.
Siyuan Ma et al. · arXiv 2608.20114 source
A Smartphone Video Builds a "Digital Twin" of a Door — Wheeled-Legged Robot Achieves 96.57% Door-Opening Success · manipulation
Opening and passing through doors is a classic long-horizon loco-manipulation task requiring both precise handle reaching and coordinated base-arm control. Video2DoorTraversal reconstructs an articulated, directly simulatable digital twin of a real door from just an RGB video, then uses a closed-loop simulation agent to convert articulation parameters into a parameterized skill program, iteratively refining failed episodes to generate physically executable demonstrations, and finally trains a dual-depth policy called ArticuACT. Perception and reasoning run entirely onboard; average success rate across five real doors was 96.57%, with 80.95% zero-shot success on unseen doors of similar structure, and a full approach-open-pass sequence takes about 13 seconds.
Xincheng Tang et al. · arXiv 2608.20251 source
SafeBranch: Letting Agents Learn Safety From Their Own Mistakes · vla
VLM-based embodied agents often violate safety constraints while completing tasks, and the difficulty is that safety-relevant signals appear at only a tiny fraction of critical steps within a trajectory. SafeBranch rolls back an unsafe episode generated by the agent itself to the step where the violation occurred, requests a safe alternative action at that point, and constructs "differ-only-at-this-step" branch pairs for alignment training — no critic is needed at deployment. On IS-Bench, SafetyALFRED, and out-of-distribution variants with unseen tasks/objects, the number of safe successes was roughly ten times that of an untrained baseline, with no loss in task success rate.
Hyunse Lee et al. · arXiv 2608.19729 source
MILD: Training Bipeds to Walk on Ground That "Gives Way" · locomotion
Deformable terrain such as sand, rubble, and planetary regolith is critical for disaster response and planetary exploration, but existing simulators fail to capture the spatiotemporal non-uniformity of such substrates. MILD introduces a physics-based discrete element method contact solver to simulate position-varying foot-ground interaction, then trains a terrain-aware controller using deep reinforcement learning with latent-variable modulation and proprioceptive estimation. Hardware experiments show the system can identify and adapt online to a wide range of ground stiffness.
Zeren Luo et al. · arXiv 2608.19955 source
Other papers today: CoToGrasp generates dexterous grasps conditioned on specified contact topology, avoiding expensive object-annotated datasets (arXiv 2608.19776 source); GOAG exploits the geometric consistency between gripper and object at the contact point to enable object-agnostic grasp planning (arXiv 2608.19759 source); LF-GICP uses a voxel normal-localizability field for tuning-free, degeneracy-aware LiDAR odometry that generalizes across four sensor types in scenes like tunnels without re-tuning (arXiv 2608.19522 source); CAViAR introduces a video dataset for traffic-accident causal reasoning, finding that current VLMs can recognize the scene but fail to correctly attribute fault (arXiv 2608.19380 source); a survey on end-to-end autonomous driving traces the architectural and evaluation evolution from behavior cloning to world-model planners and VLA systems (arXiv 2608.20111 source); a hierarchical reinforcement-learning pipeline lets quadrupeds perform explosive movements to pass through narrow gaps in confined spaces (arXiv 2608.19977 source); Inter-X++ fills a gap in human-human interaction datasets by adding dexterous hand gestures and multimodal annotations (arXiv 2608.20312 source); DreamHand uses a video diffusion model to recover heavily occluded 3D hand trajectories from first-person video (arXiv 2608.20308 source); SCAPE predicts a policy's real-world performance in a specific scenario using a small number of paired sim-real samples, reducing testing cost (arXiv 2608.19425 source).
Open Source · Tools · Benchmarks
· HiFi-UMI-2K: Simple AI released a 2,000-hour handheld-collected manipulation dataset (CC BY 4.0), drawn from a larger corpus of over 20,000 hours, 4.32 million episodes, and 480-plus scenes; the companion HiFi-UMI collection system uses head-mounted offline binocular-inertial SLAM, achieving 3mm end-effector precision without external motion capture — the team reports that 4,000 hours of pretraining reduces action error by 41% on 10 unseen tasks and improves real-robot success rate by 18.1 percentage points (technical report arXiv 2607.25895 source) source
· Franka Panda ROS 2 stack: an open-source software stack restores ROS 2 support for the Franka Emika Panda and fixes its long-standing unreliable external position-control interface, by decoupling real-time communication from the ROS 2 control loop via an asynchronous hardware interface (arXiv 2608.19740 source)
· TIER IV × Renesas: the two companies are jointly developing an open-source AI-native computing platform for autonomous vehicles source
Funding and Deals
Hyundai Motor Group → SoftBank | Sale of entire 47.5% RAI stake | Amount undisclosed · embodied
According to Hyundai Motor Group's August 21 half-year report, the group has decided to divest its entire 47.5% stake in the Robotics and AI Institute (RAI), with the shares reclassified as assets held for sale; South Korean outlet Asia Economy reports the buyer is SoftBank. RAI is a research institute spun out of the Boston Dynamics organization, focused on foundational physical AI capabilities in perception, judgment, learning, and mobility; as of the first half of last year it held roughly KRW 503.5 billion in assets and posted a net loss of KRW 140.4 billion in the same period. Hyundai's stated rationale is to "focus on Boston Dynamics as the core" of its physical AI market response, concentrating resources on product development and commercialization. Industry observers widely link this move to a possible Boston Dynamics IPO — shedding a basic-research unit unlikely to generate near-term revenue clarifies the path to profitability, though it also transfers the source of long-term technical reserves elsewhere. ⚠️ Single-source claimSource: Asia Economy source
Robonet (Hangzhou) | Series B | Nearly RMB 100 million | Post-money valuation RMB 1 billion · autonomy
Led by Hainan Jichen (affiliated with Inner Mongolia Huiming Earthwork), with multiple industrial investors including Inner Mongolia Dongfang Earthwork and Shuangcheng Earthwork participating. The company develops an L4-level, factory-installed full-stack autonomous driving solution for open-pit mine transport, with a technical approach of "strong vehicle, weak cloud" — pushing key perception, decision-making, and control down to the vehicle edge to handle mine sites' weak-network or no-network conditions. This round's investors are almost entirely mining-industry capital, reflecting a shift in the financing logic for autonomous mine transport from financial investment toward "buyers investing directly." Source: PEdaily source
Siemens Healthineers | ARPA-H funding | Up to $31.1 million | Up to $36.5 million including cost-share · adjacent
The U.S. ARPA-H agency, under its Autonomous Interventions and Robotics (AIR) program, has awarded Siemens Healthineers up to $31.1 million over five years to develop an endovascular robot capable of performing remote, autonomous mechanical thrombectomy for acute ischemic stroke; Siemens Healthineers is contributing an additional $5.4 million in cost-share. Stryker is participating as a subcontractor, extending a neurovascular interventional robotics collaboration the two companies established in 2025. Source: pharmaphorum
Global physical AI sector | H1 funding | Nearly $47.3 billion | Up nearly 80% year-over-year · adjacent
According to Crunchbase data, physical AI companies globally raised nearly $47.3 billion combined in the first half of 2026, up nearly 80% year-over-year — capital is shifting from behind-the-screen applications toward real three-dimensional environments like factories and recycling plants. ⚠️ Aggregated dataSource: Crunchbase News source
Commercialization and Deployment
49 Chinese Central SOEs Jointly Bring Procurement Lists to WRC for the First Time; State Grid's RMB 6.8 Billion Embodied-AI Equipment Procurement Enters Supplier Matching · industrial
49 Chinese central state-owned enterprises appeared collectively as a delegation at the World Robot Conference for the first time, bringing 12 categories of high-risk, complex application scenarios — inside nuclear reactor containment, 10,000-meter deep-sea environments, ultra-high-voltage power lines, high-temperature furnaces — conditions humans can barely access, which are exactly where these SOEs' robots are focused. Unlike previous years, this delegation arrived with concrete procurement demands: following the "2026 Embodied Intelligence Development Plan" set in April, State Grid announced on the same day that its RMB 6.8 billion embodied-AI equipment procurement has entered the supplier-matching stage, targeting roughly 8,500 units, including 500 humanoid live-line maintenance robots. The conference's theme also shifted on its second day from "launches" to "procurement," with the agenda focused on supply-demand matching and order fulfillment. A note of caution: conversion losses can occur at every step — from signing to delivery, from delivery to stable operation, from stable operation to repeat purchase — so whether this actually materializes still depends on field performance. Source: SmartHey source
60% of H1 Humanoid Shipments Went to Performance and Research Use; Under 20% Entered Core Production · humanoid
Following yesterday's total shipment figures, Counterpoint Research's breakdown gives a clearer picture of where the industry currently stands: of global humanoid robot shipments in H1 2026, 33.6% went to entertainment/performance and 27% to data generation and research, together over 60%; service/guiding accounted for 19%, smart manufacturing 12.8%, and warehousing/logistics 4.9%, together under 40%. In other words, under 20% actually entered core production settings. This breakdown explains why "repeat orders" was a recurring question at this year's WRC — a substantial portion of the near-300% shipment growth is not generating sustainable second purchases.Source: SmartHey, citing Counterpoint Research data source
Galbot (Chinese embodied-AI startup) × JD.com's Fully Autonomous Forward Warehouse Completes 100 Orders in One Day, Zero Manual Intervention · embodied
The forward-warehouse solution jointly showcased by Galbot and JD.com performs grasping, bagging, sealing, and organizing entirely without human intervention; on-site sorting throughput was about 1,200 items per hour, with a peak test of 2,000, and the system has completed over 100 real fulfillment orders in a single day with zero manual takeover; the solution is already running routinely at more than a dozen logistics centers nationwide in partnership with SF Express and China Post. Founder and CEO Yao Maoqing (Wang Xingxing's counterpart at Galbot) [remove — unverified] said the company has been delivering at a "thousand-unit" scale this year, and deliberately distinguishes this figure from ordinary hardware sales — counting only "model-driven production units." His framing: "customers aren't buying robot hardware, they're buying the production work the robot can complete." ⚠️ Vendor claimSource: BigGo Finance source, SmartHey source
Waymo Fully Opens in Houston, Serving Over 100,000 Riders Since February · autonomy
Waymo has ended its phased rollout in Houston, opening the service to all users — riders can now hail via the app without a waitlist. The service launched in February of this year to waitlisted users only; Waymo says it has served over 100,000 riders since then, including significant World Cup-related travel to and from stadiums. This marks another completion of Waymo's standard two-stage rollout template — waitlist first, then full opening.Source: CleanTechnica source
Aurora: At Least 200 Driverless Trucks on the Road by Year-End, Production Capacity Already Fully Contracted · autonomy
At a retail-investor town hall, Aurora (NASDAQ: AUR) said it expects at least 200 commercially operating fully driverless trucks on public roads by the end of 2026; CFO Dave Maday said production capacity to support this target is already fully contracted, with manufacturing partner Roush expected to reach a weekly output of 20 trucks (roughly 1,000 per year) starting in October. Current progress stands at 20–25 trucks by the end of Q3, with 10 approved driverless routes and 12 routes currently operating. CEO Chris Urmson emphasized that AI progress helps but doesn't solve everything — "there's a big gap between vibe-coding a demo and industrializing an 80,000-pound vehicle for highway conditions" — and said Aurora is currently the only company operating driverless trucks on public roads. The system is expected to handle light snow and cold temperatures by year-end. ⚠️ Company projectionSource: MarketBeat source
Robot.com Signs Seven-Year Deal With Sodexo, Its Largest Enterprise Deployment to Date · adjacent
Autonomous delivery company Robot.com signed a seven-year commercial agreement with food service and facilities management group Sodexo, expanding a partnership that began in 2021 on select North American college campuses into a broader scope, adding a robot advertising business as well. The company says this is its largest single enterprise deployment. Chief Commercial Officer J. Kim Fennell said the five years of experience taught both companies a lot about deployment methods and economic models, and now they're set to scale up significantly. For campus delivery, a multi-year contract matters more than any single rollout — it converts robots from a hardware purchase into an annually billed service commitment.Source: PR Newswire source
Industry Developments
Waymo Reveals Its First Self-Developed Robotaxi Chip: TSMC 5nm, Over 1000 TOPS, Already on the Road in Ojai · autonomy
In an August 20 blog post, Waymo publicly disclosed its self-developed onboard compute chip for the first time — an ASIC built on TSMC's 5nm process, delivering over 1000 TOPS of compute, already deployed in its latest-generation robotaxi, Ojai. The chip doesn't handle vehicle-wide compute; instead it sits at the front of the pipeline, processing raw sensor data and running machine-learning models — extracting information in real time from LiDAR, radar, and camera streams, including temporal denoising to improve low-light perception, and processing high-fidelity data from 13 high-resolution cameras simultaneously. The 1000 TOPS figure is close to Nvidia's stated 1000 INT8 TOPS for DRIVE AGX Thor, but the two aren't directly comparable — Waymo hasn't disclosed its own precision format, and Thor is a central SoC that also runs driving, cockpit, and infotainment workloads together. Notably, Waymo positions this chip as additive rather than a replacement: beyond TSMC, AMD, Micron, Nvidia, Samsung, SanDisk, and Socionext remain on its partner list. The company also said onboard raw compute has expanded 20x over the past eight years, with the overall system designed around three principles — responsiveness, vibration/temperature resilience, and dual-path redundancy — running two independent compute units in parallel so one can seamlessly take over if the other fails.Source: Automotive World source
Nevada Approves 8,000 Vehicles for Commercial Robotaxi Operation in One Go, Turning Las Vegas Into a Three-Way Contest · autonomy
Following an earlier episode where Tesla applied for 5,000 vehicles but regulators actually approved only 10, the Nevada Transportation Authority unanimously approved three permits this Thursday allowing Tesla, Uber, and Waymo to operate commercial Robotaxi services in Clark County, including Las Vegas — together permitting up to 8,000 vehicles over the next 12 months: Tesla capped at 5,000, Waymo and Uber each at 1,000, with Uber operating through partnerships with Hyundai-owned Motional and Zoox (Zoox separately holds a 100-vehicle network operator permit). Whether these numbers actually get deployed is another matter — Tesla Cybercab chief engineer Eric Early told the hearing directly that "5,000 has always been a ceiling," and said he'd be satisfied reaching around 2,500 vehicles a year from now. Local limousine operator associations and taxi companies explicitly opposed the permits, citing oversaturation of the commercial transport industry and road congestion. Source: TechCrunch source
China's Market Regulator: Embodied Intelligent Robots Added to Phase Two of Quality Supply-Chain Initiative, Rolling Out 166 Tasks · industrial
At an August 21 briefing, Zhang Leilei, deputy director of the Quality Development Bureau at China's State Administration for Market Regulation, said the ten flagship projects of phase two of the "quality supply-chain strengthening" initiative continue to cover priority areas such as high-end instrumentation, new energy, and industrial robots, while expanding into new areas including embodied intelligent robots and solid-state power batteries, and adding industrial mother machines (precision machine tools), drones, and advanced synthetic biomanufacturing. Phase two coordinates the full chain of standards, metrology, inspection and testing, and certification/accreditation, rolling out 166 tasks; on standards, the priority is advancing international standard development for agentic AI, brain-computer interfaces, and hydrogen energy, among other areas. Phase one's scale offers a benchmark: its ten flagship projects completed 289 tasks, with 2,060 key projects implemented nationwide involving 260,000 supply-chain enterprises; the industrial robot project alone improved the precision of 360kg-class heavy-duty robots by 2x. Moving embodied intelligence from "industrial policy support" into "quality infrastructure development" signals that the pre-mass-production groundwork — metrology, testing methods, certification systems — is beginning to be systematically built out.Source: The Paper source
Tesla Shares Roadmap With JPMorgan: AI5 Delayed to Mid-2027, Optimus External Sales at Earliest H2 2027 · autonomy
According to Electrek, Tesla recently briefed JPMorgan at its Fremont factory on its roadmap: the AI5 chip launch has been delayed to mid-2027, with a long-term goal of bringing Cybercab's per-mile operating cost down to about $0.30; the third-generation Optimus design is complete and the supply chain largely locked in, with external commercial sales targeted for as early as H2 2027, and long-term production targets of 1 million units/year at Fremont and 10 million units/year in Texas. For context: Tesla's recently disclosed cumulative unsupervised driving mileage stands at 380,000 miles, versus Waymo's over 200 million miles. ⚠️ Company projectionSource: Digital Today source
Estun Automation H1 Net Profit RMB 161 Million, Up 23x Year-Over-Year; Overseas Revenue Up 25% · industrial
Estun Automation (002747.SZ) announced on August 21 that H1 2026 revenue was RMB 2.578 billion, up 1.14% year-over-year, with net profit attributable to shareholders of RMB 161 million, versus RMB 6.6823 million in the same period last year — up 2,314.23% year-over-year; non-recurring-adjusted net profit was RMB 70.1717 million, up 498.64% (versus a loss of RMB 17.6028 million a year earlier). The breakdown shows clear divergence: industrial robots and smart manufacturing systems revenue was RMB 2.138 billion (+2.21%), while automation core components revenue was RMB 439 million (-3.78%); revenue in China was RMB 1.641 billion, down 8.82%, while overseas/international revenue was RMB 936 million, up 25.06%. Nearly flat revenue paired with a sharp profit jump indicates this improvement stems from cost and structural factors rather than demand expansion. According to MIR DATABANK data, the company ranked first among all brands by shipment volume in China's industrial robot market in H1 2026, leading for six consecutive quarters and ranking first among Chinese brands by shipment volume for eight consecutive years.Source: Caiwen source
JD.com Bets on Robotics: RMB 10 Billion in Resources, 80 RoboBase Sites, 10 Million Hours of Data · adjacent
As WRC 2026's global strategic partner, JD.com announced its robotics strategy: RMB 10 billion in resources committed by 2028, 80 RoboBase robotics facilities to be built over the next five years, and over 10 million hours of real-world scenario data to be collected within two years. RoboBase is defined as a full-lifecycle complex covering display and delivery, maintenance, pilot assembly, data collection, and iterative upgrades, paired with after-sales support capability covering more than 100 countries worldwide; JD says this will create over 100,000 after-sales engineering jobs — the labor pool pointing to the delivery-worker cohort previously trained in batches under the "Nirvana Plan," which had signed agreements with 124 schools nationwide. On August 20, JD Industrials also launched a "Robot Component Industry Development Alliance" together with Schaeffler, Qingdao Yunlu, and Yifei Intelligent, and is pushing a "standardized battery" initiative to unify battery dimensions, interfaces, and communication protocols across the industry, already partnering with more than 20 brands. Channel access, after-sales service, and data — all three levers point to the same goal: bringing the infrastructure that integrators most lack under JD's own control. ⚠️ Company projectionSource: OFweek source
UBTech Partners With BASiC Semiconductor: Silicon Carbide Into Robots, Robots Into Silicon Carbide Production Lines · hardware
On WRC's opening day, BASiC Semiconductor (9971.HK, Chinese SiC power device maker) and UBTech (9880.HK) signed a strategic cooperation agreement focused on energy efficiency optimization and low-power, long-endurance products for humanoid robots, advancing large-scale application of silicon carbide power devices in power management and motion control. Joint drive-motor actuation, fast battery charging, and edge compute power supply are all major energy consumers for humanoids, and systematic adoption of SiC devices directly targets battery life, a key bottleneck for commercialization. The cooperation runs both ways: BASiC Semiconductor will deploy UBTech humanoid robots on its own SiC production lines, jointly building a "humanoid robot semiconductor manufacturing application demonstration line."Source: Sohu source
Galbot Unveils Its First Bipedal Humanoid, ET1, Using Tennis as a Capability Benchmark · humanoid
Galbot (Chinese embodied-AI startup known for wheeled robots) unveiled its first bipedal humanoid, ET1, at WRC, currently being trained to play tennis, and will perform in the opening show of the second World Humanoid Robot Sports Games on August 22. CTO Wang He explained that tennis tests both "brain" and "cerebellum" simultaneously — requiring whole-body control to maintain balance and wrist precision to make racket-ball contact, while also deciding shot placement and how to score across the net. Behind ET1 is the company's self-developed AstraBrain-Agent agentic architecture for real-time multimodal interaction and motion control, letting the robot continuously observe its environment, understand verbal instructions, and plan dynamically rather than simply replaying pre-recorded motion or motion-capture data. On the show floor it also demonstrated folding laundry, washing and hanging clothes, and making the bed.Source: China Daily source
Hardware · Supply Chain
· Zhongke Lingxi (Chinese robotics firm) LG-603 Industrial-Grade Dexterous Hand: world first launch, hand dimensions 203×82×40mm, weight 510g, static payload 25kg, grasp-to-weight ratio 49:1, minimum grasp diameter 1–2mm, maximum grip force 85N — combining fine manipulation and heavy-duty gripping that previously required two separate devices into one hand; the company says it has shipped over 1,000 units, covering more than a dozen industrial sectors including chemical sampling, motor assembly, power grid maintenance, and consumer electronics, with the first batch of dexterous hands installed on a motor production line at Lu'an Leili Motor running continuously for over a year, accumulating 10,000 hours with zero failures (China's Ministry of Industry and Information Technology's "Humanoid Robot and Embodied Intelligence Standard System (2026 Edition)" requires dexterous-hand servo joint lifespans of ≥10,000 hours) ⚠️ Vendor claim source
· Gaoce Technology composite metal tendon cables: the company says it has cumulatively shipped over 5,000 units to dexterous-hand customers in China, supporting over 300 dexterous-hand installations, with orders covering nearly 40 domestic and international customers including dexterous-hand manufacturers, embodied-AI integrators, and university research institutes, and has completed a dedicated production line — tendon cables are the core transmission component for dexterous-hand and joint motion, directly determining end-effector precision and lifespan, and this batch of orders marks a shift from technical validation to volume supply ⚠️ Vendor claim source
· Moore Threads full-function GPU: founder, chairman and CEO Zhang Jianzhong said at WRC that the company's full-function GPUs now span the entire embodied-AI training pipeline — synthetic data generation, foundation-model training, post-training and reinforcement learning, and simulation evaluation/Sim2Real — with a single chip supporting AI compute, graphics rendering, and physics simulation simultaneously, eliminating the need for developers to switch between three heterogeneous platforms; the company's KUAE computing cluster is scaling from 10,000 to 100,000 GPUs, and the MT Lambda embodied-AI simulation platform launched this May provides an upper-layer policy development and high-fidelity physics simulation toolchain ⚠️ Vendor claim source
Top comments (0)