Today's Highlights
· A humanoid robot clocked 9.39 seconds in a 100m preliminary heat, faster than Usain Bolt's human world record
· Galbot (Chinese embodied AI startup)'s humanoid robot played fully autonomous tennis against human players in a livestreamed match
· UBTech recreated a client's actual production line 1:1 at the WRC show floor, with sorting throughput reaching nearly 1,100 units/hour
· AoYi Technology has three dexterous-hand orders in hand totaling roughly 6,000 units
· HKU's RoboDojo benchmark found the best model achieved only a 12.8% real-robot success rate, versus 100% for human experts
· China's humanoid robot shipments exceeded 40,000 units in H1, reportedly a 97% global share
Research Progress
Open Source · Tools · Benchmarks
· RoboDojo: a unified evaluation platform launched by the University of Hong Kong's Multimedia Lab together with nearly 20 universities (including UC Berkeley and Tsinghua), combining simulation testing and standardized real-robot testing in a single framework to assess generalization, memory, fine motor control, and long-horizon multi-step tasks. The results are stark — the best-performing model achieved 8.8% success in simulation and just 12.8% on real robots, versus 76.03% and 100% respectively for human experts. The project was co-launched by Ping Luo of HKU's School of Computing and Data Science and PhD student Tianxing Chen; Luo says it's the first Hong Kong-led benchmark unifying simulation and standardized real-robot evaluation, aiming to "move embodied AI from impressive demos to measurable, comparable, trustworthy progress." The open-source resources have been downloaded over 100,000 times on Hugging Face source
· JEPA-WAM: a latent world model released by StarSight AI (Chinese embodied AI startup) during WRC, with a base of just 0.5B parameters. Rather than generating pixel-level future video, it learns the environment's temporal dynamics on top of a frozen V-JEPA encoder. It introduces a Shared Predictor architecture that lets the same predictor handle both latent-state prediction and continuous action generation, propagating supervision signal from environment dynamics directly back into the action backbone. The approach can also be added as lightweight supervision on top of existing VLA models (e.g., π0.5); code is open-sourced source
· Rokae's "Xuanhe iNex": an open motion-control platform positioned as an industry-wide "cerebellum," handling dual-arm coordination, whole-body dynamics, and force-control bandwidth, paired with an out-of-the-box full-stack embodied AI development platform with open interfaces that let third-party large models or client algorithms plug in directly — an example of a component maker forgoing its own "brain" and focusing solely on the execution layer source
Funding & Deals
Dexmal (Chinese embodied AI foundation-model startup) | New funding round in talks | Target valuation ~RMB 20 billion (~US$3 billion) · embodied ⚠️ Reported terms, not finalized
Founder Tang Wenbin confirmed to Bloomberg during the World Robot Conference that the company's ongoing funding round is seeking a valuation of around RMB 20 billion, with negotiations still underway. The previous round was led by Alibaba and NIO Capital in late last year, with AI lab Z.AI (Zhipu) participating in an earlier round. The company builds foundation models for embodied AI, and this valuation target would place it in the top tier of Chinese embodied AI startups; however, the round has not closed, and the amount and investors are not yet finalized.Source: Bloomberg
Hyperion Robotics (Espoo, Finland) | Growth round | €6.4 million | ~€17.4 million raised to date · industrial
Co-led by Course Corrected and the European Innovation Council (EIC) Fund, with participation from Romande Energie's RE Ventures, and follow-on investment from existing backers Lifeline Ventures, Übermorgen Ventures, and PC Rettig Impact & Co. Founded in 2020, the company runs "micro-factories" that 3D-print concrete foundation components with robots — moving what would normally be poured on-site in a single pour into digitally designed, robot-printed, on-demand delivery from a controlled facility near the project site. Clients already include infrastructure operators such as National Grid (UK), Costain, and Anglian Water; this round will mainly fund the company's first factory, Forge I, in Flixborough, UK.Source: Silicon Canals source
PANDAG (Chinese lawn-robot maker) | Series A | Tens of millions of RMB · adjacent
Investment came exclusively from Yunshi Capital, with Xingchen Capital acting as financial advisor; proceeds will go toward product development, delivery capability, and overseas markets. Founded in 2021 in Chongqing, the company makes mowing robots for commercial landscaping settings, having previously raised funding from XBOT PARK (an incubator run by prominent roboticist Li Zexiang), MiraclePlus, and Vision Plus Capital, with a 2023 angel round from Li Zexiang's Tsingshuiwan Phase II venture fund. Its flagship product, the PANDAG G1, targets golf courses and parks; the company's website states it features a lidar-fusion system, can operate continuously for 6 hours, and covers up to roughly 48,600 square meters per day.Source: Robotics Outlook source
Yigong Precision | Series A | Amount undisclosed · hardware ⚠️ Compiled from public filings
Kaifu Capital participated in the investment; the information was compiled from a Tianyancha filing disclosed August 13, with the funding amount undisclosed. Founded in Shanghai in December 2017, Yigong Precision specializes in RV precision reducers for industrial robots; its Phase I facility has an annual capacity of 30,000 units, rising to 100,000 units once Phase II comes online. Reducer supply and cost are among the prerequisites for scaling up humanoid robot production, and this kind of upstream component maker has clearly been raising money faster than full-robot makers this year.Source: Stockstar source
Commercialization & Deployment
UBTech recreates a client's actual production line 1:1 at its WRC booth, running nearly ten industrial humanoids for eight-hour shifts · humanoid ⚠️ Vendor claim
Rather than lining up flagship models for show, the booth ran nearly ten Cruzr S2 and Cruzr Y1 units from the 9am opening to the 5pm close: the Y1 continuously palletized and depalletized, moving cartons and bins, while the S2 loaded and unloaded machined and sheet-metal automotive parts. According to on-site reporting by QbitAI, the machined-parts station requires two robots to jointly move oversized workpieces over 1 meter long, achieving positioning accuracy of under 1 millimeter despite variation in incoming parts, calibration drift, and workpiece deformation; the sheet-metal station runs its VLA model on-device. At a separate logistics-sorting station, the S2 picked skincare product boxes of varying sizes, colors, and stacking states from mixed bins and placed them onto Dematic's automated sorting wall, with QbitAI's rough on-site estimate putting peak average throughput at close to 1,100 units/hour; when a grasp fails, the robot autonomously recognizes the failure and retries with an adjusted strategy, with no manual intervention or pre-scripted routine. UBTech Vice President and Head of the Embodied AI and Humanoid Robotics Research Institute Jiao Jichao said that when clients evaluate vendors, the first thing they look at isn't benchmark scores but whether the hardware is stable and the solution actually works in practice; these deployments have moved past pilot training and into small-scale delivery. On the model side, the company claims its Thinker foundation model ranks first on 9 metrics among embodied-AI "brain" benchmarks under 10B parameters, and that its Thinker-WM world model tops the LIBERO leaderboard — these are vendor claims that have not been independently verified.Source: QbitAI source
AoYi Technology has three orders for thousands of dexterous hands each, totaling roughly 6,000 units · hardware
CEO Ni Hualiang told China Business Journal and other outlets during WRC that the company is currently fulfilling three orders of 1,000, 2,000, and 3,000 units respectively, totaling roughly 6,000 units, all of which are already in delivery. The number of partners has grown from a dozen or so in the first year, to over 100 in the second year, to nearly 300 in the first half of this year — but the distribution is highly uneven: 90% of clients remain at the proof-of-concept stage, while the remaining 10% account for over 90% of revenue, and half of those have reached larger-scale engineering deployment. LBX Finance separately cited the company's data showing dexterous-hand orders topped 10,000 units in H1. These figures say more about the state of the sector than headlines about dexterous hands "dropping to tens of thousands of yuan" — the number of customers actually buying at scale remains small.Source: China Business Journal source
Honglu Steel Structure signs order for 100 robot seventh axes · industrial
The company's smart equipment division signed separate deals for 50 units each with clients Huayuan and Jitu, totaling 100 robot seventh axes. A seventh axis adds a linear degree of freedom to a robot's base, letting a single robot cover a track of workstations that would otherwise require multiple robots — used in welding, material handling, and painting, with a modular design and configurable travel and payload. For a company whose core business is steel-structure manufacturing, this marks the first sizable milestone for its smart-equipment business moving from single-unit validation to batch delivery.Source: China.com source
Galbot's smart-pharmacy solution deployed across dozens of Chinese cities; "Galaxy Space Capsule" now at over 170 locations · embodied ⚠️ Vendor claim
The company says its humanoid-robot smart-pharmacy solution, built on the Galbot AI brain, has been deployed to hundreds of instant-retail warehouses across dozens of Chinese cities, running fully autonomously 24/7; its "Galaxy Space Capsule" retail units have reached over 170 locations across more than 40 Chinese cities, covering tens of thousands of product categories. At its WRC booth, the company demonstrated the full unattended-retail workflow — recognizing products, precise picking, autonomous delivery, and order completion without human intervention. These location counts and uptime figures are the company's own claims and have not been independently verified.Source: IPO Daily source
GAC-incubated Huilun Technology: nearly 50 GoMate Mini units deployed, longest project running continuously for 12 months · humanoid ⚠️ Vendor claim
Huilun Technology was incubated by GAC Group (Chinese automaker) in February this year to take over GAC's humanoid robot business (a prior funding round exceeding RMB 100 million was previously reported). The company says its flagship model, the GoMate Mini, uses a reconfigurable wheeled-legged design and has been deployed across 7 flagship projects with a total of nearly 50 units; the longest-running project has operated stably for 12 months, accumulating roughly 63,000 km of patrol distance and handling 189,000 on-site anomaly responses, while securing nearly RMB 10 million in commercial orders. In July this year it was incorporated into the China Security Association's security framework. The company plans small-batch trial production this year, with mass production targeted for 2027.Source: OFweek Robotics source
Pony.ai's Singapore robotaxi service integrates with local ride-hailing app Zig · autonomy
Following this week's launch with Uber in Zagreb, Pony.ai's Singapore robotaxi service is now available for booking through Zig, a local ride-hailing app, letting users hail rides directly through the platform. This is another instance of the company's "plug into existing local ride-hailing apps" strategy in Southeast Asia — since building a standalone app to acquire new users is costly, tapping into an established user base is the more common route for overseas robotaxi expansion.Source: TechRepublic source
Industry Developments
Second World Humanoid Robot Games opens; 100m preliminary heat clocks 9.39 seconds, faster than Bolt's record · humanoid
On the evening of August 22, the second World Humanoid Robot Games opened at the National Speed Skating Oval ("Ice Ribbon"), with 666 teams and 2,056 robots from 16 countries across six continents set to compete in 1,301 matches over 5 days. In the 100m preliminary heat on opening night, a humanoid robot ran 9.39 seconds, faster than Usain Bolt's 9.58-second human world record set at the 2009 Berlin World Championships, which had stood for 17 years. In earlier pre-competition testing, the Honor team's "Lightning" robot recorded 9.32 seconds with a peak speed of 14.5 m/s — this model's effective leg length was extended from 0.95m to 1.05m, and it uses a pure vision-based system for the 100m, with a camera mounted at the hip (where sway is smallest) to keep it from drifting off the track at high speed. China Daily also reported that a high jump demonstration at the opening ceremony reached 2.88 meters, versus a best of 95.64 cm at the previous Games. This year's event features 51 events in total, with all track events — the 400m, 1500m, and 4x100m relay among them — upgraded to require full autonomy, plus new combative/skill events including long jump, weightlifting, tug-of-war, and table tennis. The more notable structural shift is in event composition: 21 events are scenario-based, over 40% of the total, directly replicating real-world settings from industrial assembly and material loading to housekeeping, firefighting/rescue, and book organizing — Xinhua reports these events are designed to test robots' "last-mile" dexterity. The speed record itself is an engineering outcome pushed to the limits of motion control, structural design, and thermal management, which is a different matter from productivity; the scenario-based events over the following four days are where the real substance of this Games lies.Source: Xinhua source; Reuters source
Galbot's humanoid robot plays fully autonomous tennis against human players in a livestreamed match · embodied ⚠️ Vendor claim
At the opening ceremony's exhibition match, Galbot's humanoid robot rallied multiple shots with tennis players at the "Ice Ribbon," independently judging shot trajectories in singles play and, in doubles, observing teammate positioning to determine who should return the ball and covering the court accordingly — the entire match was broadcast live by China Media Group. Tennis is not an official event at this year's Games; this was an exhibition. According to 36Kr, this is currently the only company in the world that has brought fully autonomous tennis play to a real match setting. Behind it is the company's embodied foundation model, AstraBrain: a "brain" module reads the game state and determines shot strategy, while a "cerebellum" module coordinates the legs, torso, arms, and wrists in real time to run, balance, and swing. The training approach bypassed motion-capture replication — even fully captured human motion data can't be directly transferred, since joint structure and range of motion differ — instead placing "imperfect" human motion data into a simulated tennis environment for multi-agent training against itself; founder and CTO Wang He said the robot has accumulated training time in simulation equivalent to a human playing nonstop, day and night, for decades. When the company released AstraBrain Latent this past March, it published a video of the robot rallying against a real player; former Tesla AI lead Andrej Karpathy said at the time that on first viewing he suspected the footage was AI-generated. The live broadcast match is a harder test, since real incoming shots don't follow a training script.Source: 36Kr source
China's H1 humanoid robot shipments reported over 40,000 units, nearly double another industry estimate · humanoid
According to the "2026 Humanoid Robot Industry Development Report," China's humanoid robot shipments exceeded 40,000 units in the first half of 2026, with its share of global shipments rising further to 97%; Zhang Feng, Party Secretary of the China Institute of Electronics council and chair of the China Humanoid Robot and Embodied AI 100-Person Council, said at the event that humanoid robots are progressing from moving on stage and running in competitions to being used in homes and working in factories. Worth flagging: another recent industry estimate put global humanoid shipments in H1 at over 22,000 units — nearly half the figure above. The boundaries of what counts as "humanoid" (wheeled or quadrupedal variants included or not) and what's being measured (units off the line, delivered, or shipped) still aren't standardized, so it's worth checking whose count and which measure any given figure refers to.Source: Xinhua source
Component makers collectively unveil full robots at WRC, moving booths from the periphery to center stage · hardware
The pattern from previous years — full-robot makers occupying the prime floor space while component suppliers were tucked into the corners — has been upended this year: makers of dexterous hands, joint modules, and force-controlled arms have moved their booths to the center of the hall, and rather than just displaying gears and motors, many are now showing off complete robots of their own. Rokae gave its central booth position to a wheeled dual-arm robot, the "Star H1," relegating its force-controlled joint "Xuanji" and force-controlled arm "Taibai" to a side position; Benmo Technology (Chinese robotics component maker) split its exhibit into a components division and a robot-body division, showing three complete robots: Xingtian, TITA, and D1. The technical justification vendors give is that component specs form a closed loop under lab conditions, and only integrating them into a full robot running in real operating conditions provides a real feedback loop. But the pressure to capture more of the value chain is just as real — industry observers told LBX Finance that a complete humanoid unit sells for anywhere from hundreds of thousands to over a million RMB, while core components typically account for under 30% of that value and still face price pressure from full-robot makers; being able to offer complete robot solutions lets a supplier position itself as a systems integrator facing the end customer, meaningfully raising both deal size and pricing leverage. Most players' response has been to forgo developing their own "brain" and focus on owning the "cerebellum," while plugging in third-party models for semantic understanding and task decision-making.Source: LBX Finance source
JD Industrials leads formation of a robot component industry alliance · adjacent
JD Industrials, together with supply-chain partners including Schaeffler, Qingdao Yunlu, Suzhou Yuankong, Yeefung Intelligent, Xiamen Hongfa, Dahuan Technology, Shuangyuan Technology, and Shanghai Junyi, launched the "Robot Component Industry Development Alliance" during WRC. The idea is to let real industrial use cases define next-generation component performance targets, lifespan standards, and safety specifications, connecting the full chain from transmission, motors, and sensing to electronic control, end effectors, robot integration, and precision inspection. A JD Industrials representative's view is that "the bottleneck in going from lab to real-world deployment isn't just algorithms and models — it's core components," pointing to widespread industry problems of fragmented technology, inconsistent standards, insufficient validation, and high costs. The alliance's three-year goal is to help 100 robot makers cut costs across product lines and help 100 component suppliers double their revenue — an aspirational target, but component standardization genuinely is the practical bottleneck behind "impressive prototype, hard to scale."Source: Sohu source
Black Sesame Technologies brings automotive-grade compute to robotics, launches three-tier 48–700 TOPS compute platform · hardware
The company, which built its business on autonomous-driving chips and listed on the Hong Kong Stock Exchange in 2024 (2533.HK), showed three modules of its SesameX embodied AI compute platform at WRC: Kalos, offering up to 48 TOPS for low-speed wheeled use cases like delivery, cleaning, and inspection; Aura, with 70 TOPS of heterogeneous compute for multi-legged robots and collaborative arms; and Liora, offering up to 700 TOPS and supporting world models and end-to-end control, targeted at humanoids and multi-robot coordination — the on-site demo of on-device multimodal VLA ran at over 10 Hz. At the top of the software stack, X-Safety provides four-domain isolation to ASIL-D automotive safety standards. Chief Marketing Officer Yang Yuxin said "autonomous driving is an early form of embodied AI" — autonomous driving has 4 degrees of freedom, while robots have more degrees of freedom and more complex tasks, but the two share the same underlying technology chain. On the financial side, the company reported 2025 revenue of RMB 822 million, up 73.4% year-over-year, of which embodied AI solutions revenue was RMB 96.3 million.Source: China Business Network source
Tesla sets September 3 date for Cybercab unveiling in Austin · autonomy ⚠️ Planned event, not yet confirmed details
Tesla has announced an event on September 3 in Austin to unveil the Cybercab, a driverless vehicle designed for its robotaxi service, joining a fleet currently dominated by the Model Y. The company has already conducted real-vehicle testing on public roads and private roads at its Austin headquarters, and completed emergency-response training for local first responders in July. Production plans, the operating model, and a launch timeline will only be revealed at the event — until then, this remains just an event announcement.Source: DoNews source
North America Q2 robot orders: 8,940 units, $622 million, non-automotive customers make up 56% · industrial
According to the Association for Advancing Automation (A3), North American companies ordered 8,940 robots worth $622 million in Q2, up 4.3% year-over-year in unit terms; H1 totals reached 17,995 units worth $1.166 billion, up 2.0% in units and 6.6% in order value. The structural shift is more notable than the totals: automotive OEM orders fell 25% year-over-year, with growth in semiconductors, auto components, food, metals, and life sciences making up the gap, and non-automotive customers now accounting for 56% of Q2 orders. North American industrial robot demand is shifting from a single automotive pillar toward diversified industries — a very different trajectory from China's humanoid-focused narrative.Source: Automation.com source
Hardware · Supply Chain
· Schaeffler · molded harmonic reducer: completed large-scale validation testing, with plans to begin mass production in Germany in 2027 before gradually expanding to other regions. Using a molding process in place of conventional machining cuts forming time for key components from minutes to seconds, reducing manufacturing cost by over 25% and material consumption by over 75%; actuators account for roughly half of a humanoid robot's total manufacturing cost, so cheaper reducers flow directly into the full-robot bill of materials source
· Lens Technology × Honor: the "Lightning" robot body that ran 9.32 seconds was jointly developed by the two companies, with 132 core metal structural components covering key motion units including the head, arms, hips, and legs; its liquid-cooling system is derived from Honor's own smartphone thermal-management technology source
· Qualcomm: updated its robotics platform to add 5G connectivity and edge compute, targeting perception, motion planning, and real-time decision-making for autonomous mobile robots and humanoids source
Top comments (0)