Today's Highlights
· DiDi's next-generation Robotaxi R2 launches fully driverless paid trial rides in Beijing and Guangzhou
· Galbot (Chinese embodied-AI startup) opens three fully autonomous robot retail stores in Hong Kong, doors open September 1
· FedEx scales up Dexterity's truck-loading robots to Hagerstown hub operation
· AgiBot (Chinese humanoid startup) open-sources AGIBOT WORLD phase three, first batch of 11,430 real-robot reinforcement-learning trajectories
· Zhongda Leader (Chinese reducer maker)'s profit eaten by price war, net profit attributable to parent down 40.09% year-on-year
· Universal Robots' parent company sues Jaka (Chinese cobot maker)'s German subsidiary at the Unified Patent Court in Europe
· Momenta posts RMB 1.60 billion in first-half revenue, adjusted net loss narrows 97%
Paper Progress
VLAct: continued pretraining for VLA through representation, not more data · vla
Robot trajectory collection costs cap dataset scale, so this paper treats representation quality as the bottleneck instead, running multi-embodiment continued pretraining under a fixed data budget before task fine-tuning. It reaches 82.6% on LIBERO-Plus and 92.5% on RoboTwin 2.0, surpassing industrial VLA systems such as ABot-M0 and LingBot-VLA; on the unseen humanoid embodiment RoboCasa-GR1, it beats GR00T-N1.6 trained on the full dataset while using only 20% of the downstream trajectories. The whole pipeline uses open-source data, and training ran on just 16 GPUs. Hugging Face community interest: 49↑.
Senqiao Yang et al. · arXiv 2608.27550 source
How large a video diffusion model can 5,500 hours of driving data support · world-model
Driving data is too costly to buy or scrape at scale, which forces scaling laws to be recalculated. valeo.ai trained an entire family of models from scratch, ranging from 1M to 9B parameters, and found that validation loss follows a power law with respect to both model size and training exposure, but degrades much faster with exposure — meaning that with limited compute, extending training is more cost-effective than scaling up the model. The resulting 9B model sets a new open-source SOTA for driving video generation on nuScenes; code and weights have been released.
Victor Besnier et al. (valeo.ai) · arXiv 2608.28404 source
MAGP: giving geometric reconstruction a real-world scale · perception
Existing reconstruction methods only produce relative geometry, with scale drifting across scenes, viewpoints, and input configurations — object sizes and spatial distances don't match the real metric scale that robot actions operate in. MAGP uses scale-equivariant augmentation to push the model to recover metric geometry from camera parameters and depth observations, cutting absolute error from 2.01 meters to 0.07 meters on ETH3D, MegaDepth, and ScanNet++. Plugged into multiple robot policies, it boosts RoboTwin performance by up to 6.26%.
Fengjun Zhong et al. · arXiv 2608.27497 source
Aero Hand Open: a tendon-driven hand open-sourced together with its simulation model · manipulation
Tendon-driven hands are cheap — moving the motors out of the joints allows for a smaller hand where a single cable can drive several joints — but the tradeoff is that simulation is hard to build and coupled joints can't be commanded independently. This hand writes the cable transmission itself into the simulation model, paired with an identified bidirectional actuation mapping (including three-way coupling for the thumb) and an RL training package. Policies trained purely in simulation transfer directly to the real hand, with no fine-tuning or state estimation needed. The mechanical design, simulation, mapping, training environment, and deployment stack are all released.
Nan Wang et al. · arXiv 2608.28578 source
Walking while seated: a humanoid robot learns omnidirectional locomotion on a wheeled chair · locomotion
A near-direct-drive humanoid standing upright must continuously output joint torque, whereas sitting shifts the body weight onto a chair. The authors have the robot learn omnidirectional locomotion on a passively rolling chair, with the pelvis and seat left unfixed, propelled by intermittent foot pushes against the ground. The policy uses no motion-imitation reward — the actor only takes proprioception and velocity commands as input — and the best seated policy actually outperforms a standing policy on velocity tracking, transferring zero-shot to a real Unitree G1.
Kango Yanagida et al. · arXiv 2608.28090 source
Using contact rewards to guide exploration, a quadruped mobile manipulation robot learns to move a chair · manipulation
Non-grasping manipulation like pushing, nudging, and prying gets stuck on hybrid dynamics and sparse contact. This paper adds a dedicated exploration critic to multi-critic RL, using a dense "find contact" reward to steer the end-effector toward meaningful contact points, then gradually decaying its weight back to the task-optimal policy; candidate contact points come directly from a general-purpose grasping algorithm, which lets the approach generalize across object shapes. Across three tasks — pushing a box, moving a chair, and opening a dishwasher — the chair-moving policy was deployed on a real quadruped mobile manipulation robot.
Simone Tolomei et al. · arXiv 2608.28140 source
CoCoBench: breaking apart multi-agent collaboration failures · benchmark
Overall success rate masks duplicated effort, sequencing violations, resource contention, and handoff desynchronization all at once. CoCoBench uses 897 oracle-verified household task instances to score four collaboration components separately: task allocation, sequencing, mutual exclusion, and handoff. Across 11 mainstream MLLMs evaluated, collaboration ability turns out to be highly component-specific: a model that leads on overall score can still fail badly on one particular collaboration dimension.
Yang Chen et al. · arXiv 2608.28266 source
What happens when a robot mishears a single word · benchmark
The risk introduced by voice interfaces is rarely tested in isolation. The authors overlay simulated ASR recognition errors onto SafeAgentBench and POEX and find that some errors preserve semantic structure while amplifying harmful ambiguity, and others directly weaken the model's refusal behavior, letting unsafe plans get generated and executed. Automatic error correction sometimes suppresses the risk, but not reliably.
Sihan Jia, Oliver Lemon · arXiv 2608.28518 source
Other papers today: DeicticVLA unifies language and deictic gesture instructions into a single VLA (arXiv 2608.28108 source); CAVE-NAV does 3D autonomous navigation of underwater caves with a VLM (arXiv 2608.27793 source); ChainSplat learns deformation dynamics of cables, ropes, and hoses from multi-view RGB video (arXiv 2608.28570 source); PanelShield adds verifiable closed-loop safety planning for industrial panel operation (arXiv 2608.28305 source); a year of running through subarctic forests, with the authors reviewing the practical difficulties of autonomous navigation under canopy cover (arXiv 2608.27628 source); Picking Bins Empty uses online self-learned grasp points to solve deadlocks in emptying industrial bins (arXiv 2608.28175 source); AcrossVAM1.0 splits robot video prediction into object-level motion and dense appearance streams (arXiv 2608.28491 source); PHR-VLA adds planning-horizon reasoning to VLA (arXiv 2608.27609 source).
Open Source · Tools · Benchmarks
· AGIBOT WORLD phase three: AgiBot has released a reinforcement-learning dataset, with a first batch of 11,430 real-robot interaction trajectories covering industrial and household scenes across 14 categories of interaction tasks, including plugging in network cables and unlocking doors with a key. The data was collected by the next-generation Spirit G2 platform and will be released in five phases, with digital-twin simulation data released in parallel. Real-robot RL trajectories have long been scarce in public datasets, and this batch is large enough to support third-party reproduction. source
· Microduck: a $399 bipedal robot duck jointly released by Hugging Face and its French subsidiary Pollen Robotics — 25 cm tall, under 800 grams, packing 15 servos, a wide-angle camera, a small lidar, and dual IMUs. The full-stack open-source release is the selling point: 4,096 virtual ducks can be trained in parallel in MuJoCo for RL, producing a new gait in roughly one to two hours, then exported as ONNX and deployed to the physical unit. The device is manufactured by Shenzhen Seeed Studio (Chinese hardware contract manufacturer). source
· Berkeley Humanoid Lite: an open-source humanoid reference design from UC Berkeley — 0.8 meters tall, about 16 kg, with full hardware priced under $5,000 at US market rates. The actuators use brushless DC motors paired with 3D-printed cycloidal reducers and position encoders; the housing and gears are designed to fit within a 200×200×200 mm desktop-printer build volume, with the larger 6512 actuator listed at about $188. source
Funding & Deals
NVIDIA × Hugging Face｜Acquisition｜$12.9 billion · adjacent ⚠️ unconfirmed report
The Information reported on August 27 that NVIDIA has agreed to acquire the open-source model hosting platform Hugging Face, with CNBC, Reuters, and Forbes following up subsequently; neither company has officially confirmed the deal, and reports say a final agreement has not been signed and could still change. On the embodied-AI side, Hugging Face acquired France's Pollen Robotics last year and just released Microduck on August 27, so ownership of the LeRobot ecosystem would change hands along with the deal. If completed, this would be NVIDIA's largest acquisition ever, roughly twice the scale of its 2020 acquisition of Mellanox.Source: The Information / CNBC
Shunxing Qidian (Hangzhou)｜Seed round｜tens of millions of RMB · hardware ⚠️ compiled from public filings
Investors include two industry players, Galbot (Chinese embodied-AI robotics company) and Zhongke Diwuji, alongside Bamboo Capital (Chinese VC), Hangzhou Kechuang Group, Zhaofeng Co., and Duowei Capital. The company positions itself as a one-stop service provider for embodied robotics, handling the full chain for OEMs from core components and joint modules to full mass-production delivery, without building its own consumer-facing brand. As integrator OEMs have piled in, contract manufacturing and mass-production delivery are starting to attract dedicated funding of their own. Data compiled by Securities Star based on Tianyancha's public disclosure filed August 30.Source: Sina Finance source
ABLE Labs (South Korea)｜Series A｜about $8.7 million (KRW 12.1 billion) · adjacent
Led by Innopolis Partners, with participation from the Korea Development Bank, Capstone Partners, Horizon Investment, Magna Investment, Futureplay, and NAU IB Capital. The company builds lab automation robots and plans to expand from liquid-handling robots into autonomous laboratories, while also using the experimental data infrastructure it builds to train models. 2025 revenue was KRW 1.29 billion, up 113% year-on-year, and first-half 2026 order backlog already exceeds full-year 2025 revenue by more than double; customers include KAIST, Pusan National University, and the Korea Research Institute of Chemical Technology, and the company is also co-building an organoid drug evaluation platform with Johns Hopkins University.Source: WOWTALE source
Xuanchuang Robotics｜Series A｜tens of millions of RMB · embodied
Following the closing of a Series A1 round, this round's investors have been revealed as Shenzhen Qianhai Fangzhou Capital, Guangyang Co., and others. The company recently won an order for more than 100 units from PetroChina, with deliveries running at 20 units per month.Source: DoNews source
Commercialization & Deployment
DiDi's next-generation Robotaxi R2 launches fully driverless paid trial rides in Beijing and Guangzhou · autonomy
R2 was co-developed with GAC Aion and purpose-built for Robotaxi service, running DiDi Autonomous Driving's full-stack L4 software and hardware, equipped with 33 sensors and a tri-domain fused central computing platform, plus multiple layers of redundancy, with safety standards aligned to C-NCAP and Euro NCAP five-star ratings. After delivery in January this year, the vehicle first ran public-road testing in Guangzhou, Beijing, Shenzhen, and elsewhere; it can now be booked via the DiDi app in designated demonstration zones in the two cities. The cabin features a 17.3-inch ceiling-mounted screen, and voice commands alone can complete PIN verification, start the ride, adjust the air conditioning, and open windows. Just the day before, Caocao Mobility pushed RoboX into Abu Dhabi — Chinese Robotaxi fleets are simultaneously moving toward removing safety drivers and toward overseas markets.Source: PR Newswire source
Galbot opens three fully autonomous robot retail stores in Hong Kong, doors open September 1 · embodied
One store each in Hung Hom, Kai Tak, and Wan Chai — Galbot's first stores outside mainland China. Founder and CTO Wang He said at the launch ceremony that this step moves the company from R&D into actual operations and is a key step toward global markets, and that he hopes to build embodied-AI labs jointly with Hong Kong universities and local industry partners. Galbot has already opened roughly 200 stores across nearly 50 cities in mainland China. Financial Secretary Paul Chan attended and cited Hong Kong's dense population, diverse retail environment, and high degree of internationalization; Hong Kong Investment Corporation has been a Galbot shareholder since 2024. What retail stores demand of robots isn't motion difficulty but continuous availability while unattended — three stores opening on the same day is about density, not a single showcase.Source: China Daily source
FedEx scales Dexterity's truck-loading robots from pilot into the Hagerstown hub · industrial
The deployment pairs Dexterity's Foresight world model with its Mech loading robot, at the Hagerstown Hub in Maryland, at a scale considerably larger than the earlier pilot. Mech is a dual-arm system built narrow enough to work inside trailers; Foresight combines vision, depth, and touch to predict how actions affect the physical world, and loading requires weighing space utilization, load stability, and speed across three spatial dimensions plus time. Trailer loading is one of the most physically demanding steps in package logistics that traditional automation has long failed to crack, and FedEx loads tens of thousands of trailers a day across the US. Kawal Preet, FedEx's Executive Vice President of Planning, Engineering, and Transformation, said the expanded partnership is about "evaluating and deploying physical AI in real-world environments." Dexterity founder and CEO Samir Menon called Hagerstown a template for replicating across the network.Source: AI Insider source
Momenta posts RMB 1.60 billion in first-half revenue, adjusted net loss narrows 97% · autonomy
The first half-year report since its Hong Kong listing: revenue of RMB 1.60 billion, up 76% year-on-year, gross profit excluding share-based compensation of RMB 1.21 billion at a 75% gross margin, with adjusted net loss narrowing to RMB 14.097 million. Revenue splits into two segments — technology development services at RMB 995 million (62.1%) and licensing services at RMB 607 million (37.9%); licensing revenue was only about RMB 23 million back in 2023. As of end-June, the company had 105 mass-produced vehicle models and 219 secured design wins cumulatively, with vehicles running its system now exceeding 1.10 million units. R&D spending was RMB 1.16 billion, or 72.6% of revenue. Robotaxi is already being tested in Shanghai, Suzhou, Abu Dhabi, and Munich, and in July the company obtained Germany's KBA nationwide urban road L4 testing permit; Momenta told National Business Daily that its first mass-produced Robotaxi model will launch in Q4 2026, running the R7 world model, with several hundred vehicles deployed domestically and internationally by year-end. With L2 penetration already past 70% and continuing to spread into lower-tier markets, whether the 75% gross margin can withstand the price war is the real thing to watch in the second half.Source: National Business Daily source
MEgo Technology's 20,000th MEgo capture device rolls off the line, embodiment-free data surpasses 1 million hours · adjacent
The 20,000th MEgo unit was delivered by founder and CEO Yao Maoqing to JD.com, going directly into use in real business scenarios; the same day, MEgo announced an embodiment-free data business partnership with Tencent Robotics X Lab. The company says cumulative embodiment-free data production has surpassed 1 million hours, all collected in real open environments, covering 22 major scene categories, over 10,000 real environments, over 50,000 object categories, and more than 500 fine-grained tasks. MEgo (Chinese motion-capture-data startup) was incubated inside AgiBot before spinning off independently, and has raised hundreds of millions of RMB, with investors including China Telecom, HongShan (formerly Sequoia China), and Baidu Ventures. Now that capture devices have scaled to 20,000 units, the competitive front on the data-supply side is shifting from device count to data structure and scene coverage.Source: The Beijing News source
Hai Robotics' HaiClimber surpasses 10,000 units contracted worldwide · industrial ⚠️ vendor-reported
Since its global launch in 2025, HaiPick Climb has been selected for customer projects in 12 countries across Asia, Europe, North America, South America, and Australia, with customers including Anta, Panasonic, Arvato, Itochu, Mettler Toledo, and True Protein. Founder and CEO Rocky Chen said what's more notable than the 10,000-unit figure is that 8 customers have expanded existing systems or repeat-purchased for new projects. Two different framings appear in the same release: the body text says "contracted for use in customer projects," while Americas Sales VP Hunter Senn's quote says "deployed" — and there is typically a delivery lag of several months between a contract and installation.Source: Construction Business News ME source
Industry Trends
Universal Robots' parent company sues Jaka at the Unified Patent Court in Europe · industrial
Teradyne Robotics has sued the German subsidiary of Jaka Robotics at the Unified Patent Court (UPC) in Copenhagen, alleging infringement of both software and hardware patents, covering a broad range of collaborative robot models that Jaka sells in the EU. The patents at issue are held by the Danish subsidiary Teradyne A/S, tied to the Universal Robots A/S business unit. Teradyne Robotics Group President Jean-Pierre Hathout said the company would enforce its patent and IP rights against what it considers "unlawful and improper imitation." This is Teradyne Robotics' second IP lawsuit in Europe in 2026, following an earlier copyright suit in Germany over Universal Robots software against another robotics maker's subsidiary. The ruling can apply across 17 of the UPC's current 18 member states, with non-UPC countries such as the UK and Spain also drawn into the proceedings. This is the first time Chinese cobot makers' expanding European market share has run into a pan-European patent enforcement tool.Source: All About Industries source
Zhongda Leader's half-year report: revenue up 15.6%, net profit attributable to parent down 40% · hardware
This manufacturer based in Cixi, Zhejiang, mass-produces all three types of reducers — planetary, RV, and harmonic. First-half revenue was RMB 597 million (+15.61%), net profit attributable to parent RMB 27.78 million (-40.09%), non-recurring-adjusted net profit RMB 19.51 million (-48.15%), with overall gross margin at 23.28%, down 4.46 percentage points year-on-year, and operating cash flow turning negative to -RMB 56.49 million. The revenue mix has shifted: smart actuator units brought in RMB 253 million (42.40%), becoming the largest revenue source for the first time; precision reducers brought in RMB 134 million (22.39%), also the single biggest drag on profit this period. Per pricing cited by Robot Vision Forum (Chinese robotics industry outlet) from public industry information, harmonic reducer unit prices have fallen from over RMB 3,000 during the era of import monopoly to the low-thousands range, with some models down near RMB 800; China's harmonic reducer annual production capacity is heading toward a million units, while global humanoid shipments in the first half were still only in the tens of thousands, with each robot using 8 to 14 harmonic reducers. The same report notes that Leaderdrive (Chinese harmonic reducer maker) still holds gross margin above 30% on the strength of high-end products and top-tier design wins. Humanoid shipment volumes ramping up and component makers actually turning a profit are, for now, two separate things.Source: Robot Vision Forum source
Li Auto unveils next-generation autonomous driving architecture MindVLA-o1 at GTC · autonomy
Zhan Kun, head of foundation models, offered the diagnosis that today's end-to-end systems are all essentially "learning to drive by watching 2D video" — BEV flattens the world and loses height, while OCC is three-dimensional but lacks semantics. MindVLA-o1 is built around a native multimodal MoE Transformer, using a native 3D ViT encoder that unifies 3D geometric and semantic understanding in a single encoding pass, reducing lidar to a "ruler" that geometrically calibrates vision. Li Xiang (Li Auto founder and CEO) said unified modeling stably extends perception and reasoning to over 500 meters. On the behavior side, VLA-MoE with parallel decoding is paired with discrete diffusion for multi-round trajectory refinement; on the training side, step-by-step reconstruction has been replaced with feed-forward scene reconstruction, nearly doubling rendering speed and cutting overall training cost by about 75%. 3D ViT is compute-hungry, and deployment relies on Li Auto's own Mach 100 chip at 1,280 TOPS per chip, first fitted to the all-new L9. The team evaluated nearly 2,000 model architecture configurations.Source: Autohome source
Qiyuan and Zhijie both disclose L3 approval progress the same day, both still awaiting ministry-level review · autonomy ⚠️ vendor-reported
Xuan Wei, product general manager at Qiyuan Auto, said on social media that the Qiyuan GT7 already meets the design standards of China's national L3-level autonomous driving architecture and has completed L3 access testing and filing, with further rollout still pending review by national ministries and related regulations. He described the L3 threshold as requiring that the eight systems — power supply, compute, perception, positioning, communication, steering, braking, and interaction — must never all go offline simultaneously in the event of component failure, and the vehicle must be able to safely pull over. Zhijie's framing goes one step further: product manager Li Xinyue said in a livestream that the Zhijie RX has been undergoing L3-related autonomous driving testing in Shanghai, Wuhu, and elsewhere, though aggregator headlines rendered this as "access testing," which differs from the body text. Richard Yu (Huawei executive) said the Zhijie RX is the first coupe SUV built for L3, with L3-architecture-equipped orders accounting for over 90% of pre-sales at an average price of RMB 380,000. Both companies remain at the filing-and-testing stage, and whether the feature can be activated depends on regulators, not the vehicles themselves.Source: PChome source, Sina Auto source
BYD and PaXini upgrade from an investment relationship to strategic cooperation, opening vehicle plants to robots · embodied ⚠️ vendor-reported
The agreement was signed August 25 at BYD's global headquarters in Shenzhen and announced by BYD on the 31st, covering both data collection and robotics. BYD contributes complex, exacting vehicle- and component-manufacturing processes, extensive factory environments, and supply-chain resources, giving robots access to real industrial test settings and generating operational data usable for training embodied models; PaXini (Chinese tactile-sensing robotics startup) contributes tactile sensors, multi-dimensional perception, and multimodal embodied data collection. Founded in Shenzhen in 2021, the company builds high-precision tactile sensors, dexterous hands, and the Tora series of humanoids; it raised over RMB 1 billion in March at a valuation above RMB 10 billion, with JD.com also among its shareholders. BYD invested in PaXini in April 2025 and had previously invested in AgiBot. PaXini has previously partnered with Geely, with its Tora-One performing wire-harness assembly and glue application at the Qiantang plant; the company's stated figures at the time were 8 hours of continuous fault-free operation and a 99.8% task success rate — figures for a single demonstration run, not the same thing as sustained production-line operation.Source: CnEVPost source
California Teamsters sue over the legislative process behind new driverless trucking rules · autonomy
California's driverless trucking framework, drafted since 2021, took effect April 28, requiring 1 million miles of testing before commercial freight operation. Teamsters California filed suit in Alameda County Superior Court on August 5, not challenging the substance of the rule but only the process — alleging that the DMV pushed the rule through a fast-track process meant for regulations with first-year costs or benefits under $50 million, "evading the law's requirement that the agency study and publicly disclose the economic impact." The lawsuit hasn't frozen approvals — the permit queue continues to move, with 28 permit holders including Aurora, Kodiak AI, Gatik, and Plus. Plus's self-published readiness scorecard on August 10 showed a 93.4% safety case, 99.8% autonomous-driving mileage, and 85.2% of trips with no remote assistance, with a target of removing the driver by 2027. Its commercial pilot isn't in California but on a 600-mile daily route along the I-35 corridor in Texas, run jointly with International Motors and Ryder. What's constraining these companies for now is the calendar, not the courts.Source: FreightWaves source
Chinese robotic lawnmower makers turn en masse toward Europe · adjacent
According to IDC data, global robotic lawnmower shipments grew 63.8% year-on-year in 2025 to 1.992 million units, with the top six all Chinese manufacturers. Ninebot, Roborock, and a cluster of startups are preparing for IFA Berlin, opening September 4. Shenzhen-based LDROBOT (Anthbot brand) moved from being a lidar supplier into making complete units; co-founder and CEO Gavin Guo told Nikkei that lawnmower revenue more than doubled year-on-year to RMB 200 million in the first half of this year, now accounting for 38% of total revenue, with a target of reaching the top three within three years. Also Shenzhen-based Mova (TerraMow brand) uses a triple-camera approach, bypassing lidar and GPS to bring prices down to €669–€1,149; founder Harry Zhang, a DJI alumnus, says his reasoning is that users skew older and won't tolerate configuring a device via smartphone. There is almost no lawn-care demand within China itself, so this product line has been a pure export business from day one, and US import restrictions have pushed essentially all the growth toward Europe.Source: Nikkei Asia source
GXO's CEO: warehouses need better robot hands, not humanoids that can run and jump · industrial
GXO, a logistics company with $13.6 billion in annual revenue, is running pilots with five humanoid robot suppliers and will add a European pilot this year. CEO Patrick Kelleher said the record-setting sprints and high jumps at the World Humanoid Robot Games in Beijing were of no use to him — what he wants is a robot that can unload trucks and also sort lipstick. He's focused on what he calls the "single-digit problem," meaning getting robot hands closer to human hands: "One important advance we've been involved in this year is getting a humanoid's hand to move like this, which requires multiple flexible joints and will change the range of tasks a humanoid can do." Industrial robots typically have four to six degrees of freedom in their motions, whereas humanoids can now exceed 100. He stressed this isn't about replacing the existing 150,000 employees: "Our warehouse environments have 25% turnover, and we can't hire enough people as it is." GXO is simultaneously developing fully autonomous forklifts and humanoids that can work in cold-storage facilities at minus 32 degrees.Source: Fortune source
Hardware · Supply Chain
· Qualcomm Dragonwing IQ10: Anshuman Saxena, Qualcomm VP and GM of ADAS and Robotics, said the platform has shipped to multiple customers since its launch earlier this year, with several Dragonwing-based products now in mass production in China, spanning AMRs to humanoids; publicly disclosed partners include Figure AI and NEURA Robotics, and the company has also set up a robotics center in Japan. He pushed back on the "single big brain" approach, arguing intelligence should be distributed to wherever the actuators are. source
· XenseVision Prima 1: a purely direct-drive dexterous hand with 22 active degrees of freedom and five-finger visuo-tactile sensing; at a trade show it drew building outlines with a paintbrush, adjusting grip force in real time based on friction between the brush and paper, aimed mainly at academic research and model training. Alongside it, the Flex 2 is a 23-DOF hybrid-drive variant with air cooling, aimed at precision assembly and industrial sorting. Founder and CEO Xia Yuxuan said the two lines are complementary, not a replacement for one another. source
· Robot Component Industry Development Alliance: launched by JD Industrials during WRC 2026, with Schaeffler Greater China e-mobility division president Chen Xiangbin attending the launch ceremony. The alliance spans transmission, motors, sensing, motor control, end effectors, whole-machine integration, and precision testing, aiming to use real industrial applications to set lifespan and safety specifications for next-generation components. source
Top comments (0)