Today's Highlights
· Global humanoid robot shipments topped 22,000 units in H1, up nearly 300% YoY, with Zhiyuan Robotics (Chinese humanoid startup) alone accounting for 40%
· Unitree fell 18.70% on its second trading day, erasing about RMB 167 billion in market cap from the intraday high on day one
· Unitree also launched a seven-axis bionic arm the same day, priced from RMB 9,900, with repeatability of about 0.1mm
· Baidu's Apollo Go went live on Uber — in Dubai, users hailing an UberX may now be matched with a fully driverless car
· Pony.ai × Uber launched in Zagreb, the first Uber-hailable robotaxi service in Europe
· Waymo's next-gen Ojai rolled out fleet-wide across three cities, with roughly 300 units currently in the commercial fleet
· GEN-1.5 treats a 3–12 second demo as a prompt, averaging 59% success across ten previously unlearned tasks
· GigaBrain-WBC-0.5 achieves a 99.3% fall-recovery rate, 16.8x the strongest baseline
Research Progress
GigaBrain-WBC-0.5: Letting Whole-Body Control Policies "Know What They Can Do Next" · world-model
Whole-body motion tracking policies have previously trained only on flat, empty scenes, never learning how terrain and object contact reshape their own dynamics — an approach that relies on ever-expanding reference motion libraries and breaks down once feasible behavior changes with the environment. This work proposes the first behavior world model for humanoid whole-body control: a causal transformer that jointly predicts the next action, next state, and a distribution over next latent behavior commands. At deployment, this distribution is used online to identify infeasible commands and pull them back toward learned behaviors, letting the robot execute in a "best-effort" manner. It leads across four scenario categories against three large-scale tracking baselines — 81.3% success on terrain interaction (4.3x the strongest baseline), 83.1% under infeasible commands, and 99.3% fall recovery (16.8x the strongest baseline); on real hardware, weights trained on the Unitree G1 transfer to the Maker L01 with simple fine-tuning.
Ziyang Cheng et al. · arXiv 2608.18234 source
ADEPT: Learn to "Set Things Straight" First, Then Learn Everything Hard · manipulation
High-DOF multi-finger hands have long needed costly from-scratch exploration for every new task. ADEPT first does large-scale reinforcement-learning pretraining on a general object-rearrangement task, then treats this behavior as a prior to post-train downstream policies, using a recipe of behavior-cloning distillation, critic warm-start, and conservative on-policy updates to address the rapid degradation of pretrained capability during fine-tuning. The resulting distilled perception-student policy achieves zero-shot sim2real on two embodiments: a 23-DOF Kuka-Allegro setup (dual RGB) and a 29-DOF Flexiv-Sharpa setup (dual RGB plus five-channel visuotactile sensing), completing long-horizon tasks from difficult initial states at near-human speed.
Jayjun Lee et al. · arXiv 2608.19182 source
Redoing Push-T With a Coding Agent: Zero Demonstrations, 100% Success · manipulation
Push-T is a landmark imitation-learning benchmark, and Ken Goldberg's team took a different approach this time: letting an LLM coding agent (Claude Code running Fable 5) write its own algorithmic solution, with no human demonstrations at all. The agent found an online 2D gym simulation on its own, worked out the pushing dynamics through simulated experiments, and iteratively optimized, ultimately reaching 100% success with 46% fewer steps than the best diffusion policy trained on 200 human demonstrations. It also self-generated a curriculum extending the task from T to the entire alphabet (Push-A through Push-Z), and generated 3D cross-embodiment simulation code with visual feedback for both Franka and UR5 arms. The significance of these results lies not in Push-T itself but in what it suggests about which manipulation tasks may not actually require data-driven policies.
Shuangyu Xie et al. (UC Berkeley) · arXiv 2608.18227 source
SoftVTBench: Completing the Task Doesn't Mean You Didn't Crush the Object · benchmark
Evaluation of deformable-object manipulation has long looked only at task success, even though a policy can succeed while slipping or over-squeezing the object entirely. This dataset provides 4,000 expert demonstrations across 50+ assets (including volumetrically deformable objects and visually identical rigid twins), with 20Hz synchronized multi-view RGB, dual-finger tactile images and marker-point motion, proprioception, language, and actions, plus finite-element ground truth reserved for evaluators only. From this, the authors define a "deformation-aware success rate," DSR — counting a trial as successful only if the task completes AND peak normalized deformation stays within tolerance. Across Diffusion Policy, π0.5, and FastWAM, all 12 in-distribution configurations exhibited "successes" that violated the deformation tolerance, accounting for 0.7%–24% of each configuration's successes.
Bowen Jing et al. · arXiv 2608.18701 source · Signal: HF↑12
The "Embodiment Gap" in Robot Foundation Models: Generalizing Isn't the Same as Running on Your Robot · vla
This survey gives a name to a problem currently obscured by the scaling narrative: the gap between reusable models, representations, or data and actually executing on a target robot — the embodiment gap. Using two axes, "type of shared structure" and "stage requiring adaptation," the authors map existing approaches, covering shared semantics and perception, shared data and interfaces, and learning cross-embodiment correspondence. They also propose a reporting standard for adaptation effort — because reporting only success rates hides exactly how much work remains when switching to a new robot.
Yukiyasu Domae et al. · arXiv 2608.18433 source
GS-VLA: Move the Camera a Little, VLA Success Drops From 90% to 10% · vla
Existing VLAs implicitly assume identical camera configuration between training and deployment. The authors found that a small shift in camera mounting can, in the worst case, drop LIBERO success from about 90% to about 10%. The work reframes viewpoint shift as a local novel-view synthesis problem — under the assumption that perturbations are confined to a small neighborhood of the workspace, viewpoint normalization reduces to a scene- and policy-agnostic disocclusion task. A 4-million-parameter 3D Gaussian canonicalizer is attached in front of a frozen VLA, recovering much of the lost performance across multiple architectures, unseen task sets, and different perturbation scales without touching the policy weights.
Yechan Park et al. · arXiv 2608.19066 source
DA-WAM: A Driving World Model Must Not Just "Predict Accurately" — It Must "Predict Usefully" · autonomy
Existing approaches decouple future representation learning from planning optimization, or have multiple candidate trajectories share the same predicted state, diluting the action-tied consequences that should guide selection. DA-WAM unifies predictive representation learning, action-conditioned future modeling, and trajectory scoring under a single decision objective: an action-conditioned predictor generates a distinct future latent state for each candidate trajectory, which is then evaluated by a future-latent-conditioned factorized scorer. Predicted latent states for expert-matching trajectories are supervised by observed future representations, while safety-critical hard negatives receive additional supervision near planning boundaries. Achieves SOTA on NAVSIM-v1/v2.
Ruiguo Zhong et al. · arXiv 2608.19085 source
Other papers today: RoboEdit rewrites human manipulation videos into action-consistent robot videos via an automated pipeline producing 174K aligned video pairs across seven embodiments (arXiv 2608.18948 source); Dream2Reward learns a "successful latent displacement field" from positive demonstrations only, penalizing wrong direction, overshoot, and stalling, yielding dense reward without failure labels (arXiv 2608.18787 source); LabDex targets dexterous manipulation in chemistry labs, organized across atomic skills, composite tasks, and long-horizon experiments — the first platform unifying real hardware and simulation (arXiv 2608.18618 source); GelSight Mini tactile sensor's 3D force-map estimation transfers zero-shot across units, with image-reconstruction SSIM of 0.9338 (arXiv 2608.18240 source); PartialBiGrasp infers hidden geometry from partial viewpoints for bimanual grasping (arXiv 2608.19188 source); LT-Mem uses volatility-aware spatiotemporal memory for lifelong scene understanding (arXiv 2608.19059 source); CL4D does language–4D contrastive pretraining to add geometric depth reasoning for dynamic scenes (arXiv 2608.18734 source); USR-Drive jointly denoises 3D Gaussians and detection boxes into a unified driving-scene representation (arXiv 2608.19036 source).
Open Source · Tools · Benchmarks
· Noitom HiPHI: Noitom announced at WRC 2026 it is freely opening its humanoid robot training data to embodied-AI developers source
· Unitree R1: officials say they will open-source the control software and interfaces for this seven-axis robotic arm; the dev environment is based on Ubuntu, with XT30 and CAN485 support source
Funding & Deals
Digital Fifth Era (Hangzhou, Chinese robotics firm) | Series A1 + A2 | Over RMB 1 billion cumulative · embodied
The A1 round was jointly invested by Xigao Investment, Bank of China's BOC AIC Fund, and Sanfeng Investment; the A2 round was jointly invested by Xinneng Venture Capital, BOC AIC Fund, Zhongshan Venture Capital, Hongruida Investment, Boyan Technology, and Jinchuang Investment. Founded in 2024, the company builds an "embodied brain," launching its ultra-few-shot FAM series of embodied manipulation foundation models. Its self-developed wheeled embodied robot stands 189cm tall with 28 whole-body degrees of freedom; the company says it has deployed in logistics, flexible industrial production lines, and energy-inspection scenarios, and has already secured orders worth several hundred million RMB from overseas/international markets. State-owned AIC funds backing two consecutive rounds is an increasingly common funding structure this year among embodied-foundation-model financings.Source: PEdaily source
Veeda Innovation (Toronto, Canada) | First round | Over $90 million · world-model ⚠️ single-source account
Founder Sanja Fidler was formerly NVIDIA's VP of AI Research; the company was registered a month before she left in July. Co-founders Zan Gojcic and Huan Ling are also NVIDIA alumni. Investors are Khosla Ventures and Radical Ventures, with offices in Toronto, Zurich, Singapore, and Mountain View. The company builds world models that simulate physical reality, calling its simulator "the Matrix for physical AI" — a direction of travel shared with fellow Toronto company Waabi, which uses simulation to replace expensive and safety-risky real-world training. The funding figure comes from people familiar with the matter; Veeda declined to comment.Source: The Globe and Mail source
Wonik Robotics (South Korea) | Proposed raise | Over KRW 300 billion (about $217 million) · hardware ⚠️ rumored account
South Korea's National Growth Fund is reportedly advancing a major direct investment in this specialized robotic-hand maker; if approved, it would be the fund's first robotics investment. The policy fund's total planned size is KRW 200 trillion over five years. Sources in the investment banking industry say the deal may use a mezzanine structure such as convertible bonds or redeemable convertible preferred shares, primarily via capital increase rather than secondary share transfer, with specific terms still pending the investment review committee's decision; government officials say the proposal remains under review. Wonik Robotics is unlisted, with Wonik Holdings owning a 96.37% stake; it is co-developing the next-generation Allegro Hand with Meta, and the raise is earmarked for building its own data center to accrue data generated by robotic hands and industrial robots.Source: bloomingbit source
Agility Robotics (US) | SPAC listing advances | Pre-money valuation $2.5 billion · humanoid
Following June's announced merger with Churchill Capital to go public, the company this week split its long-combined CFO/COO role: Michael Beer, who joined in July, becomes CFO focused on the listing, while Jennifer Hunter, who previously held both roles, moves to focus solely on operations and production ramp-up. The deal is expected to yield over $600 million in total proceeds, with $421 million from Churchill's trust account and about $200 million from a PIPE. Beer says the goal is to complete the deal within the year, with some market reports suggesting a listing as early as next month. The company's customers include Amazon and Toyota's Canadian manufacturing operations.Source: CFO Dive source
Current Robotics | Seed through Pre-A, three rounds | Several hundred million RMB cumulative · embodied
First public disclosure of its funding progress. Financial investors include BV Baidu Ventures, Hillhouse Ventures, Oasis Capital, Monolith Capital, Qianhai Fangzhou, and Fosun Chuangfu, with strategic investors including Zhiyuan Robotics, Galbot, and Jike Technology. Funds are directed toward large-scale human whole-body data collection and toward whole-body dexterous manipulation foundation models and interactive world models. The lineup of strategic investors — all embodiment and foundation-model companies — suggests this round is buying a position on the data-supply side.Source: Guandian.cn source
Commercialization & Deployment
Baidu's Apollo Go Goes Live on Uber, Fully Driverless Robotaxis Open for Hailing in Dubai · autonomy
Starting August 20, users in parts of Dubai's Umm Suqeim and Jumeirah neighborhoods who hail an UberX or Uber Comfort may be matched with a fully driverless Apollo Go vehicle, or can select the "Autonomous" option directly in the app to increase their odds — no second app required, which is the most substantive change for riders in this rollout. This marks the first time the Baidu–Uber partnership has actually opened to riders. The fleet is operated by New Horizon Luxury Transport, using the sixth-generation RT6 — an all-electric vehicle seating up to three people with more than 30 sensors. Uber's global head of autonomous driving, Sarfraz Maredia, called it the first public-road validation of its "multi-partner autonomous network"; Yang Nan, head of Baidu's international autonomous-driving business, called Dubai its first overseas city running both self-operated and partnership models simultaneously. The two companies say Apollo Go now operates in 28 cities globally with over 350 million kilometers of cumulative autonomous driving, including over 240 million kilometers fully driverless.Source: Gulf News source
Pony.ai × Uber × Verne Launch in Zagreb, Europe's First Uber-Hailable Robotaxi · autonomy
Following last week's announcement that Pony.ai would enter five European cities with Uber, planning to deploy over 2,000 vehicles, the first city has launched: on August 19 in Zagreb, Croatia. Pony.ai supplies the autonomous-driving technology, Croatian mobility company Verne manages the fleet, and Uber provides the hailing interface, letting users hail an autonomous vehicle directly in the Uber app. Each vehicle still carries a licensed safety operator, making this Europe's first commercial robotaxi service bookable through a mainstream ride-hailing platform — though it's still a step short of removing the safety operator. Compared against the same-day Dubai launch, the two paths Chinese autonomous-driving companies are taking overseas/internationally — operating independently versus riding on an existing platform — were both validated on two continents this week.Source: Uber Investor Relations, Silicon Republic
Waymo's Next-Gen Ojai Rolls Out Fleet-Wide Across Los Angeles, Phoenix, and San Francisco · autonomy
All riders hailing a car in the three cities may now be matched with an Ojai; once fleet scale grows, riders will be able to choose between an Ojai and the older Jaguar I-Pace. A Waymo spokesperson says the commercial fleet currently has about 300 Ojai units, with deployment planned this year to Denver, Las Vegas, and San Diego. Ojai carries the sixth-generation autonomous driving system and the Gemini in-car assistant; the vehicle itself is a Geely-owned Zeekr MPV built on the SEA-M platform, designed to be cheap to build, cheap to repair, and durable under near-continuous operation — though under current US trade policy, tariffs on imported Chinese-built vehicles raise the landed cost of each unit. Bloomberg reported the same day that Waymo has developed custom chips for its robotaxis. For a company betting on profitability through per-vehicle cost reduction, a fleet across 11 cities with only 300 Ojai units suggests the ramp-up of the cost-reduced model is only just beginning.Source: TechCrunch source
Tesla Robotaxi Drives Through Barrier Posts in Austin, Continues On · autonomy ⚠️ compiled account
A passenger who frequently rides Tesla's driverless robotaxi filmed the entire episode from the front passenger seat: the vehicle stopped repeatedly in front of a row of plastic barrier posts, inched forward and backed up several times, and ultimately drove through the posts into a closed-off area before continuing on normally — with no safety operator in the driver's seat. Online commenters, comparing the footage against street-view imagery of the buildings visible in the video, concluded the posts had been in that location since at least February 2024. The repeated stopping suggests the system detected and responded to the obstacle, meaning the problem more likely lies in path planning and decision logic — though the video alone cannot confirm the cause, nor whether remote operator intervention was involved. Worth noting for comparison: at the Q2 earnings call on July 22, Tesla VP of AI Ashok Elluswamy said Robotaxi had completed over 380,000 miles of unsupervised driving across six cities in two states "with no notable incidents," calling the safety record "impeccable." Whether this incident meets NHTSA's reporting threshold remains unconfirmed.Source: kkj.cn source
RealMan Plans Nearly 1,000 RealBOT Deployments This Year, Feeding Teleoperation Data Back In · embodied ⚠️ company-stated plan
The company unveiled its GLN (Global Link Network) real-world deployment plan at WRC 2026, saying it will deploy nearly 1,000 RealBOT units within its own network in 2026, feeding perception–manipulation–decision data generated by remote operation back into training to improve cross-environment generalization. Live demonstrations included medication retrieval and restocking at a smart pharmacy, power-distribution-room inspection, collaborating with a chef on pastries, and remotely operating equipment at its Changzhou embodied-AI factory; the company says GLN maintains millisecond-level latency even over thousands of miles. Both the deployment figures and latency claims are company-stated, with no third-party verification; still, using teleoperation as a data-collection channel echoes the same bet Current Robotics is making with whole-body data collection, covered in this issue's funding section.Source: Interesting Engineering source
LG CNS Deploys Quadruped Security Robots at Seoul's Tower Palace · industrial
South Korea's LG CNS announced deployment of quadruped security robots at Tower Palace, an upscale residential complex in Seoul's Gangnam district, with full operations beginning in Q4 for night patrol and anomaly detection. Quadruped robots' commercial use cases have historically concentrated on power, chemical, and factory inspection; moving into routine residential security is a newer step, one that will require handling unstructured foot traffic and residents' privacy boundaries.Source: The Korea Herald source
Industry Developments
Global Humanoid Shipments Topped 22,000 Units in H1 — But Two Authoritative Counts Differ by Nearly Double · humanoid
Counterpoint Research's humanoid robotics research service says global humanoid robot shipments exceeded 22,000 units in H1 2026, up nearly 300% YoY. Zhiyuan Robotics led with about 9,700 units and over 43% market share, advancing its A/G/X product lines evenly; the X2 spawned multiple commercial variants for entertainment performance, service guidance, research/education, and data production, while the G series is being deployed in clusters on consumer-electronics and automotive-parts production lines through partnerships with Longcheer Technology and Joyson Electronics. Unitree shipped over 7,000 G1 units for a 31% share, holding its research/education base steady. AgiBot (Galbot's competitor, Chinese embodied-AI startup) — wait, correction: Galaxea Robotics (Chinese embodied-AI startup) shipped over 1,100 cumulative units, about 5% share, and beyond its smart-pharmacy partnership with Meituan has already developed specialized VLA models for industrial scenarios. UBTECH delivered in the thousands, for a 4.4% share, with management raising its 2026 shipment guidance for the Walker S series from 2,000–3,000 units to 5,000 units.
The same day, Zhang Feng, Party Committee Secretary of the China Electronics Society council and chairman of the China Humanoid Robot and Embodied AI Hundred People Association, released the "2026 Humanoid Robot Industry Development Report" at WRC, stating that China's humanoid robot shipments in H1 already exceeded 40,000 units, with China's share of the global total rising further to 97%. The two figures directly conflict: by the China Electronics Society's count, China alone shipped nearly twice Counterpoint's global total. Neither source specifies in public materials which models are included in the count (whether wheeled forms, non-bipedal forms, research/education full units, or embodiment kits are counted), so the more prudent reading is to treat these as two mutually incommensurable tallies rather than average them into a single industry growth rate.Source: Counterpoint Research source, China News Service source
Unitree Falls 18.70% on Second Trading Day; Founder's First Post-IPO Speech Points to the "Last Few Centimeters" · humanoid
On August 20, Unitree opened down 6.51% at RMB 790, fell intraday to as low as RMB 685.01, and closed at RMB 687, down 18.70% for the day. Total market cap fell to RMB 277.9 billion, down about RMB 167 billion from the intraday high of RMB 444.9 billion on listing day. Behind the pullback is a gap between valuation and performance: the prospectus shows revenue of RMB 159 million, RMB 392 million, and RMB 1.7 billion for 2023–2025 respectively, with net profit of -RMB 11.15 million, RMB 95.47 million, and RMB 278 million; Q1 2026 revenue was RMB 423 million, up 68.49% YoY, but non-GAAP net profit was RMB 40.25 million, down 52.55% YoY. The company forecasts H1 non-GAAP net profit of RMB 236–283 million, down about 6.43% to 21.97% YoY.
The same day, founder Wang Xingxing gave his first public speech since the listing at WRC's main forum, saying candidly that robots' efficiency and generalization on factory floors still fall short of humans, and that the company has not yet rolled out real-world work tasks at scale. He said the industry's biggest bottleneck is insufficient generalization — a deviation of just a "few centimeters" can cause task success rates to collapse — and that reaching embodied AI's "ChatGPT moment" would optimistically take 2 to 3 years, or as long as 5 to 10 years in a slower scenario. That a company just priced by capital markets at a P/E of 219x had its founder say, on the very same day, "we haven't scaled real deployment yet" is a contrast worth noting more than the stock price itself.Source: 21jingji source
Unitree Launches R1 Seven-Axis Bionic Arm, Priced From RMB 9,900 · hardware
The R1 weighs 5.5kg; the company states a repeatability of about 0.1mm, maximum joint speed over 180 degrees/second, maximum reach of 650mm with dexterous hand attached, and rated payload of 2kg, with force feedback and collision detection. The end effector can be swapped between grippers and dexterous hands, targeting research/education, industrial assembly, and service robotics. The price is the whole weight of this news: previously, imported bionic arms with comparable DOF typically cost well over RMB 100,000, and mainstream domestic seven-axis collaborative arms in China ran RMB 30,000–80,000. At RMB 9,900, this shifts the whole price bracket down a full tier, letting budget-constrained universities and startups buy real hardware in bulk for the first time to run VLA and world-model algorithms, rather than staying confined to simulation. Worth clarifying: this is positioned for research/education and light service use, not the same category as industrial-grade arms built for 24-hour continuous heavy-duty work in terms of mean-time-to-failure, impact resistance, and lifespan.Source: TechNode source
Beijing Humanoid Robot Innovation Center Launches Unified Model and Lightweight Humanoid, Partners With Mercedes-Benz · world-model
Beijing Humanoid Robot Innovation Center launched its unified embodied-AI model Pelican-Unify at WRC, integrating VLM visual understanding, action-model motor control, and WFM world-model prediction into a single architecture, closing the loop from understanding to reasoning to rehearsal to execution, and validated end-to-end on its Tianyi hardware. The architecture is compatible with multiple embodiment types, aiming for one brain to drive multiple robot models. The center also announced commercial deployment of Pelican-VL 2.0, supporting task planning, intent understanding, and human-robot interaction. Its lightweight humanoid Tiangong Omni made its debut — standing just 1.35m and weighing 39kg overall, with onboard motor cerebellum and high-compute multimodal perception pre-integrated for edge-side large-model deployment; demonstrations included navigating plum-blossom piles, climbing stairs, and office printing tasks. The center also announced a deep strategic partnership with Mercedes-Benz, with its Tiangong 3.0 serving on-site as a "silicon-based greeter," autonomously handling welcoming, model explanations, and consultation interactions, with plans to enter Mercedes-Benz dealerships, auto shows, and delivery ceremonies going forward.Source: Gasgoo source
UBTECH Prices Ultra-Realistic Humanoid U1 Pro at RMB 169,800, Sales Start in September · humanoid ⚠️ manufacturer's account
The U1 series has three versions; the top-spec U1 Pro is priced at RMB 169,800, targeting family companionship, elder care, business reception, and museum guiding. The company states over 3 hours of language-interaction battery life on a full charge, and that it does not involve physical labor. Hardware includes 88 motion joints, a "dual-fulcrum + four-bar-linkage" bionic cervical structure, a 33-axis facial tendon network, and a bionic face driven by 19 active degrees of freedom, said to display over 300 composite micro-expressions. On the interaction side, it runs a proprietary Resonance-LM emotion model, with claimed micro-expression recognition accuracy above 90% and speech-to-lip-sync latency controlled within 20ms. The company says total orders across the U1 series had already surpassed 13,000 units as of the conference opening — both order volume and performance metrics are manufacturer-disclosed, with no third-party verification observed. Ceding "getting work done" for now to industrial robot models, and instead using mature language-interaction technology to capture the companionship market, is this product line's pragmatic choice.Source: Sina Finance source
Generalist AI Launches GEN-1.5: A 3–12 Second Demo as a "Physical Prompt" · world-model ⚠️ company's account
A 3–12 second demonstration is fed directly into the model's context as a "physical prompt," after which the robot executes the task without any further training. Across ten tests including opening jars and pulling money from a wallet, the company reports an average success rate of 59%; with five minutes of additional data and ten training steps, success rises to 83%. The model can chain two prompts into longer sequences, can use demonstrations from simulation, and can partially imitate human hand motions. The company says these capabilities emerged spontaneously during eight-plus months of pretraining on interaction data, without explicit training for them. Other teams have previously demonstrated in-context learning, but mostly limited to a few task types; Generalist claims this is the first demonstration spanning a broad range of tasks. Worth noting: all demonstrated tasks are short and simple, and all results come from the company itself with no independent replication.Source: the-decoder source
Xiaomi's Next-Gen Humanoid "Tiedan" Debuts After Four Months of Training · humanoid ⚠️ company's account
At WRC's "Embodied Garden" booth, Xiaomi's next-generation humanoid picked flowers from a flower wall, grasped incense petals in its palm to scatter, and handed them to onlookers. The robot stands 1.7m tall with 66 joints throughout its body, shorter than the roughly 1.76m first generation from 2022 — officially explained as more closely matching the height of frontline auto-factory workers. Lei Jun emphasized that actions like handing flowers, handshakes, fist bumps, and heart gestures were autonomously understood, judged, and executed by the robot via a large model, rather than remote control. The booth demonstration remains a capability showcase, and the degree of autonomy has not been independently verified.Source: Jingbao.com source, kkj.cn source
Hyundai Motor Group Bets $6.5 Billion on a "Physical AI City" · adjacent
Group Chairman Euisun Chung said at a California AI summit in late July that Hyundai is "moving beyond the traditional boundaries of automaking to accelerate its transformation into a physical AI solutions company." Its largest partner is NVIDIA, with the two collaborating on AI compute infrastructure for digital twins in auto manufacturing and for autonomous driving; the group has also committed to jointly developing humanoid robot technology with DeepMind, and announced in July it had acquired full ownership of Boston Dynamics. The centerpiece is an industrial park in Saemangeum, southwestern South Korea — the group is investing KRW 9 trillion (about $6.5 billion) starting this year to build robotics, water electrolysis, and solar facilities across 1.12 million square meters, alongside a data center equipped with about 50,000 NVIDIA GPUs, plus a city-scale physical AI training center. The group plans to produce 30,000 robots annually in the US by 2028, with the next-generation Atlas humanoid to be trialed first at the Metaplant America assembly plant in Georgia before gradual rollout.Source: Nikkei Asia source
Hardware · Supply Chain
· Octopus Power: unveiled a full-stack "brain–hand–data" system at WRC, including an embodied world model, bionic dexterous hands, and a data collection solution source
· Silicon Century (Chinese robotics company): showcased a full product matrix, running a 24-hour non-stop livestream with its dexterous hand as an endurance demonstration source
· Vadzo Imaging: launched the Armor-1335CRA-FPD3, a 13-megapixel autofocus FPD-Link III camera with coax power support, targeting AGVs and robotics source
This Week's Watch
Embodied AI Enters the International Standards Table at ITU-T for the First Time · adjacent
"Cloud-Edge-Device Collaborative Architecture: Application Scenarios, System Requirements and Overall Architecture," led by China Mobile, has been approved for standardization at ITU-T, entering the 2025–2028 standards cycle with a planned release in December 2026. The standard builds a unified technical system across five dimensions: standardizing industry-scenario requirements (covering home, education, etc.), establishing a three-tier "super-brain–brain–cerebellum" collaborative architecture, specifying cross-domain task scheduling and adaptive transmission mechanisms, defining collaboration capabilities and network-slicing scheduling requirements for gateways, optical access, and compute nodes, and building out accompanying performance evaluation, device operations, and security-compliance frameworks. This is an ITU-T international standard for collaborative communication in embodied-AI scenarios, led by a telecom operator. Coming a week after China's mandatory L3/L4 national standard was finalized (previously reported), the two developments point in the same direction: the rule-making layer for embodied AI and autonomous driving is extending from domestic Chinese standards into international standards bodies, while unified communication interfaces are exactly the prerequisite for cross-vendor, multi-robot deployment at scale.Source: Xinhuanet Zhejiang
TrendForce: China's Humanoid Robot Market at About RMB 15 Billion in 2026, Growing at Least 60% More in 2027 · humanoid
TrendForce's estimate on August 19 puts China's humanoid robot market at RMB 15 billion in 2026, with growth of at least 60% expected in 2027 as production capacity expands and use cases multiply. The firm observes that Chinese humanoid robots have begun commercial acceptance testing across automotive, 3C electronics, aviation, logistics, and energy sectors, and notes that WRC 2026 held its first-ever "Procurement Day," concentrating procurement matchmaking and supply-chain connections — seen as a signal of the industry's center of gravity shifting from technology showcases to real demand. This is a single firm's own estimate, not analyst consensus; set alongside the conflicting Counterpoint and China Electronics Society shipment figures elsewhere in this issue, it's worth noting the three sources are actually measuring three different things — "market size," "global shipments," and "China shipments." The genuinely comparable data points remain the two TrendForce flags: which products already have confirmed customers and have begun bulk delivery, and whether early customers are placing repeat orders.Source: TrendForce, TechNews
This Week's Supply Chain: The Cost-Reduction Cycle Shifts From "Full Robots" to "Standalone Actuators" · hardware
Stringing together this week's scattered signals, the clearest trend is a shift in where cost reduction is landing: the past two years saw whole-robot prices cut from over RMB 500,000 down to the RMB 30,000 range; this week the shift is toward standalone actuators being sold externally and priced down independently — Unitree's R1 seven-axis arm from RMB 9,900 (see Industry Developments above), with joints, robotic arms, and dexterous hands now sold as standard components to downstream integrators and labs. This rewrites the cost curve from "amortized internally by the OEM" to "amortized industry-wide, per component."
Pulling the timeline back to 30 days, several named figures support the same direction: per a June 25 industry survey by TMTPost, six-axis force/torque sensors dropped from RMB 20,000–30,000 last year to RMB 5,000–8,000, a decline of over 60% in a year (per Wuxi Quanzhibo). The same company's integrated joint-module overall cost fell 10%–15% YoY, with yield rising from 85% last year to over 98%, and rotary-joint monthly capacity reaching ten-thousand-unit scale. 3D vision camera entry prices fell to under RMB 1,000, down about 70% from two years ago, mainly benefiting from spillover capacity from consumer robots like robot vacuums. High-energy-density battery packs saw costs fall about 25%. On the transmission side, a representative advance is Noshok Robotics's mass production of a 1.5mm-diameter planetary roller screw (5.5mm nut diameter, C5 precision, ±0.01mm repeatability, 50N maximum load), with mass-production pricing entering the hundred-RMB range.
This week's most notable point of tension isn't in these already-falling segments, but in two areas that remain bottlenecked: precision transmission components and high-end tactile sensors. OEMs can absorb joint and motor costs through in-house R&D, but struggle to independently absorb the upstream costs of these two categories. When every company is compressing end-market prices simultaneously while upstream core-component costs haven't fallen in step, it's component suppliers' margins that get squeezed. What's worth watching next, then, isn't who cuts prices further, but real-world measurements of lifespan and maintenance cost for sub-RMB-10,000 actuators under continuous heavy-duty, high-frequency reciprocating use — the gap between low price and industrial-grade reliability doesn't yet have public data to fill it.Source: TMTPost, 36Kr
Top comments (0)