Today's Highlights
· Wayve×Uber secures London ride-hailing license, beating Waymo and Baidu to supervised Robotaxi launch
· Nvidia open-sources autonomous driving VLA Alpamayo 2 Super, 34B weights ready for direct commercial use
· Xiaomi's embodied foundation model Xiaomi-Robotics-1 officially open-sourced, trained on 100,000 hours of real robot data
· Hai Robotics' shelf-climbing robot reaches 10,000 units in global deployment
· Amazon's largest Kansai hub goes live, deploying roughly 3,000 transport robots in a single warehouse
· Unitree begins preliminary price inquiry today, estimated IPO price around RMB 104/share, valuation above RMB 40 billion
· Daegu, South Korea completes the country's first humanoid robot safety certification center, with KRW 18.7 billion invested over five years
· Tsinghua AIR×BAIR continuous-time world model generates at arbitrary frame rates and can even run inference backward
I. Research Progress
Quo Vadis, World Modeling? Where Should World Models Go Next · world-model
This position paper, trending at 27↑ on Hugging Face today, is worth reading for how it relocates the definition of "world model" entirely: classic world models predict physical state transitions, but a continuously self-improving agent actually needs "usable information" — execution outcomes, retrieved experience and skills, verification signals. Based on this, the authors propose Agent-Centric Interactive World Proxies, organizing six proxy categories by feedback modality — dynamics, space, execution, memory/experience, skills, and reward/verification — and mapping three tiers of function: in-context guidance at inference time, producing rewards and synthetic rollouts at training time, and co-evolution between agent and proxy. This is a clear course correction against the current tendency to equate world models with "video prediction."
Yu Yang et al. · arXiv 2608.02713 source
Light-Loco-Parkour: A Single Policy for Whole-Body Humanoid Parkour · locomotion
Previously, whole-body humanoid control either tracked expressive motions but was terrain-blind, or reacted to terrain in real time while leaving arms and torso idle; this work solves both sides with one deployable policy. The policy takes only onboard depth maps and velocity commands as input, deciding on its own whether to walk, balance, climb, descend stairs, or vault over obstacles — with no reference inputs, skill labels, hardcoded state machines, or runtime motion generators. Skills are learned by expanding from a small set of seed motions into terrain-matched feasible references, with successful zero-shot transfer to real hardware both indoors and outdoors.
Hongming Chen et al. · arXiv 2608.02653 source
How Should VLAs Actually Use Proprioceptive State? · vla
Nearly all VLAs consume robot proprioceptive state, but the methods vary wildly — serialized into text prompts, projected into vision-language prefixes, or fed directly to the action expert — and mostly using only the current frame, with no systematic comparison to date. Holding backbone, data, action representation, and evaluation protocol fixed, this paper implements 5 interface variants, compares them across 45 atomic tasks + 20 compound tasks, and sweeps state history length from 1 frame to 96 frames, answering three open questions: whether state actually helps, how much history helps, and where it should enter the model. These are conclusions with immediate engineering applicability.
Yiren Zhao et al. · arXiv 2608.03052 source
DRIFT: The Adversarial Robustness of Flow-Matching VLAs Is an Illusion · vla
It was previously widely believed that flow-matching VLAs like pi0 are inherently more robust to adversarial perturbations than autoregressive VLAs — this paper points out that's only because past attacks ignored the multi-step denoising ODE. The authors attach a universal adversarial patch to a robot gripper to directly attack the denoising velocity field, with a counterintuitive core finding: attacking only the first denoising step is stronger and cheaper than attacking a wider step window (stemming from gradient conflicts specific to input-space optimization, the exact opposite pattern from training-time backdoors). Across all four LIBERO suites on pi0 and pi0.5, a single small patch nearly breaks every task that was previously solvable.
Hoseong Tae, Jong-Seok Lee · arXiv 2608.03207 source
Shooting for Contact: Turning Kinematic References into Dynamically Feasible Trajectories · locomotion
Motion retargeting typically only cares about kinematic similarity, producing references that are extremely hard for RL policies to reproduce, especially for contact-rich behaviors. This paper embeds a differentiable simulator into a nonlinear program to perform contact-implicit multiple shooting, solving contact, friction, collision, self-collision, and joint limits simultaneously without needing to pre-specify contact timing. The result is faster imitation learning training and lower tracking error, and on Unitree G1 it achieved zero-shot sim-to-real transfer for contact-rich crawling and a 180-degree jump turn.
Sergio A. Esteban et al. · arXiv 2608.03116 source
Functional Safety Certification for Industrial Humanoids: Where the Fail-Passive Gap Lies · locomotion
What's holding back industrial humanoids isn't whether they can walk or grasp, but that functional safety certification for legged platforms remains immature — this paper precisely names the core problem the fail-passive gap: a legged robot's safe state is one "actively maintained by control," while the underlying assumption of ISO 13849-1 / EN 60204-1 is that de-energizing equals safe — but cutting power to a walking biped instead creates a new hazard source, falling. The authors use an externally quantifiably-certifiable safety chain (light curtains, e-stops, fail-safe PLC, wireless PROFIsafe) as a probe to localize the uncertifiable residual gap to the robot's own reaction chain, and validated feasibility on a semi-enclosed pick-and-place work cell using the Unitree G1 EDU. The authors explicitly do not claim end-to-end PL e / SIL 3 — this restraint itself signals that the barrier to humanoids entering production lines still sits on the standards side.
Caiwu Ding et al. · arXiv 2608.02809 source
ODEWorld: A Continuous-Time Embodied World Model · world-model
Mainstream video/latent-space world models are built on discrete time, understanding the world at whatever cadence the sampling rate happens to be, while real physics evolves continuously. Tsinghua AIR and UC Berkeley BAIR propose PT-Flow (Physical Time Flow), having the model learn "which direction the world is changing right now, and how fast," then integrating along the velocity field with an ODE solver to obtain the latent state at any point in time. A companion dynamics-representation decoupling hands the static background off to the initial-state memory, so the latent space only describes change; using Jacobian-vector products to project actual visual change velocity provides first-order supervision, sidestepping the JEPA approach's long-standing reliance on bolt-on regularizers to prevent representation collapse — under the RankMe protocol, its 768-dimensional dynamics latent space achieves an effective rank of 425.2, higher than the equally 768-dimensional DINOv2 CLS features (376.1) and the 1024-dimensional V-JEPA 2 (203.7). In practice, the same model can generate at any frame rate without retraining, fill in missing intermediate frames from sparse observations, and even run inference backward by reversing the integration direction.
Tsinghua University Institute for AI Industry Research (AIR) × UC Berkeley BAIR · Analysis: Tsinghua University Institute for AI Industry Research source
Bimanual Manipulation on an 8GB Budget: Real-World Deployment Testing on Entry-Level Jetson · manipulation
Imitation learning policies are almost always evaluated on workstation- or datacenter-grade GPUs, and the real cost of deploying them onto embedded hardware has barely been measured. This paper crams an entire bimanual SO-101 system onto a Jetson Orin Nano Super (8GB), using a desktop GPU only for offline training. Three findings are useful for engineering: ACT converges to usable performance within 100,000 steps (19/20 trials), while Diffusion Policy still hasn't converged after twice as many steps (0/10); after conversion to TensorRT, FP16 cuts inference latency from 114.02ms to 17.93ms, and INT8 to 12.65ms without dropping success rate; and TensorRT's generic INT8 calibration only quantizes the ResNet18 backbone, with none of the 145 transformer layers quantized — explaining why INT8 is still 28% faster than FP16 despite barely shrinking in size.
Ekansh Singh et al. · arXiv 2608.03938 source
Other papers today: UVT modifies not the VLA architecture but the supervision objective, using a unified visual-motor latent goal to jointly encode actions and scene transitions, with the largest gains under constrained training budgets (arXiv 2608.03563 source); SARF defends against "attention hijacking" from physical adversarial patches by fine-tuning only the vision encoder, cutting OpenVLA's failure rate on LIBERO from 100% to an average of 28.6% (arXiv 2608.03231 source); Track4Action distills a world-frame 3D tracker into a VLA, requiring no tracker at deployment time (arXiv 2608.03727 source); Continue or Replan? lets a VLA adaptively decide when to replan rather than on a fixed cycle (arXiv 2608.03483 source); CUDA MPC moves the entire optimization loop onto the GPU, achieving a real-time-feasible horizon one to two orders of magnitude longer than CPU solvers (arXiv 2608.03051 source); LiLa-WAM compresses a world-action model down to end-to-end trainable on a single 24GB GPU (arXiv 2608.03701 source); SLAMFormer-∞ claims to be the first geometric transformer with no explicit distance upper bound on either front-end or back-end (arXiv 2608.03429 source).
Open Source · Tools · Benchmarks
· Xiaomi-Robotics-1: Xiaomi's embodied foundation model was officially open-sourced today, covering the full pipeline from real-robot post-training to deployment along with benchmark reproduction code; Xiaomi states it was trained on over 100,000 hours of real robot manipulation data plus over 10,000 hours of cross-embodiment post-training, achieving 57.4% average success rate on RoboCasa365 and 74.5% on RoboCasa, with weights and code posted to GitHub and Hugging Face source
· OpenETA: an open-source implementation that brings the digital-agent paradigm into the physical world — the Planner selects one Tool call at a time, the Interface controls execution, and the World returns results and new observations, letting robots verify outcomes, revise plans, and turn both success and failure into reusable experience; offers a swappable Planner, composable Tools/Skills, auditable memory, and replayable trajectories (arXiv 2608.03924 source)
· Principles of Robot Autonomy: a robotic autonomy textbook compiled by Marco Pavone's team drawing on years of teaching at Stanford, bridging classical robotics and modern physical AI under a unified conceptual framework, with a Jupyter notebook and implementation-based exercises for each topic (arXiv 2608.03496 source)
· HandTok: a reproducible test bed for a hand-motion tokenizer released alongside DigitCode, organizing symbolic representations hierarchically by the hand's anatomical units, cutting quantization error by three-quarters (arXiv 2608.03127 source)
II. Funding and Deals
Moove | New Round | $250 Million | $2.1 Billion Valuation · autonomy
This company builds the "infrastructure layer" of autonomous driving mobility — vehicle assets, financing, and fleet operations, the heavy lifting that Robotaxi scaling can't avoid but that no one wants to do — and this round's funding is explicitly earmarked for expanding that layer globally. Capital in the Robotaxi space is spilling over from the algorithm side toward the asset and operations side; Uber's $1.25 billion vehicle-supply agreement with Rivian is another facet of the same trend.Source: Business Wire source
Estun Automation (002747.SZ) | Acquisition | RMB 487 Million | 100% Stake in Estun Cool-Drive · industrial
Following its mid-July disclosure of a planned all-cash acquisition, this filing sets the total deal value at RMB 487 million; upon completion, the company will indirectly hold 100% of Estun Cool-Drive and consolidate it into its financial statements. Estun states the acquisition is meant to fill out its "heavy-duty industrial robots + lightweight collaborative robots + embodied intelligence robots" full-scenario product lineup. Industrial robot-arm manufacturers using acquisitions to fill in their embodied-intelligence gap is a common path among Chinese production-line suppliers this year.Source: Sohu source
Unitree Robotics | STAR Market IPO Preliminary Price Inquiry | Targeting RMB 4.202 Billion Raised · humanoid
Following the underwriter's valuation range, August 5 is the preliminary price-inquiry day, with the issue price to be set on August 6, and online/offline subscription opening simultaneously on August 10. The public offering covers 40.4464 million new shares, 10% of total share capital post-issuance. According to CCTV Finance, the market estimates an IPO valuation above RMB 40 billion, implying an estimated issue price of roughly RMB 104/share and a subscription deposit of RMB 52,000 per lot — but this is a back-calculated estimate based on projected valuation, with different estimates ranging from RMB 40 billion to RMB 56 billion, and the actual pricing has yet to be determined. Per the prospectus, the company reported 2025 revenue of RMB 1.699 billion and net profit of RMB 278 million.⚠️ Estimated figuresSource: Sina Technology source
Xihu Digital Intelligence | Seed Round + Angel Round Series | Over RMB 100 Million Cumulative · world-model
Investors include Innoangel Fund, Oriental Jiafu, Zhengxuan Investment, Tianqi Capital, Xihu Innovation Investment Fund, and Jinma Investment. The company, incubated from Westlake University's Autonomous Intelligence Lab, was founded only this January and builds a general-purpose physics engine: pulling physical simulation out of the graphics pipeline to perform dynamics computation directly in latent space, predicting physical causality like fluid sloshing or rigid-body collisions without needing to render photorealistic imagery — the company says this cuts compute requirements by 1 to 2 orders of magnitude. Founder Yu Kaicheng previously led the scaling of visual perception algorithms at Alibaba DAMO Academy. This round's funding will go toward building a dataset of tens of millions of physical interaction trajectories.⚠️ Company's own accountSource: Stockstar source
Mariana Minerals (US) | New Round | $310 Million · industrial
Builds autonomous mining systems; this round's amount ranks among the top tier in the current mining automation space. The company's CEO framed this round explicitly in the context of the critical minerals race — the driving force behind mining automation is shifting from cost reduction to supply chain security.Source: CNBC source
Foundational Industries (US) | Seed Round | $25 Million · industrial
Aims to build "AI-native factories" — organized from the ground up around AI and robotics, rather than automation bolted onto existing factories. This is a rare from-scratch factory-building approach amid the current industrial embodied-AI wave, and the $25 million seed round also signals capital's acceptance of the underlying thesis.Source: AI Insider source
A Personal Mobility Robotics Company | Nearly $100 Million Raised to Date · adjacent
Founded by a former DJI executive, focused on personal mobility robots. Per 36Kr's exclusive report, cumulative funding has reached nearly $100 million — the spillover of drone supply chains and engineering teams into embodied AI continues, and this is already the second company with a DJI-background founding team to secure funding above the $100 million level recently.⚠️ Single-source accountSource: 36Kr source
III. Commercialization and Deployment
Wayve × Uber Secures London Ride-Hailing License, Beating Waymo and Baidu to Launch · autonomy
Wayve's autonomous vehicles have obtained a London Private Hire Vehicle (PHV) license — the same category of license held by ride-hailing operators — meaning the service can now operate on the road as a compliant commercial offering rather than remaining confined to testing. Under the current authorization, each vehicle still carries a safety operator, hands off the wheel but ready to intervene when needed; fully driverless operation would require a different license and further approval. Users who have signed up can begin trying it this summer, ahead of full rollout. The launch uses Wayve's development vehicle, a Ford Mustang Mach-E, equipped with cameras and radar; Wayve says the future fully driverless version will be based on a Nissan-manufactured model. This move clarifies the three-way race in London: Waymo is already testing on London streets and plans to launch a driverless version this year via its own app; Baidu's Apollo Go began testing last week and expects to operate publicly through Freenow by 2027. The timing is especially significant for Uber — its exclusive US partnership with Waymo expires in 2028, and Uber has already stated it plans to offer autonomous mobility with partners including Wayve and Nuro across more than 10 cities worldwide this year.Source: The Business Times source
Hai Robotics' Shelf-Climbing Robot Reaches 10,000 Units in Global Deployment · industrial
On August 3, Hai Robotics (Chinese warehouse robotics maker) announced its shelf-climbing robot has reached 10,000 units in global deployment. The shelf-climbing design is a warehouse robot category where the system uses only this single robot model, paired with shelving, totes, workstations, charging stations, and the HaiQ management platform in a "5+1" configuration, requiring no positioning codes to be deployed — plug-and-play, keeping implementation costs below traditional solutions. Customers span footwear/apparel, fast-moving consumer goods, retail, beauty, pharma, and cross-border third-party logistics, including Anta, Li-Ning, Panasonic Logistics, Itochu Logistics, and Zongteng Cangchu overseas warehouses. Reaching 10,000-unit scale with a single robot model is rare, solid evidence of real scaled deployment in warehouse robotics.Source: Guandian.cn source
Amazon's Largest Kansai Logistics Hub Goes Live, Deploying Roughly 3,000 Transport Robots in a Single Warehouse · industrial
The Mukogawa FC in Amagasaki, Hyogo Prefecture, which opened on August 5, is Amazon's largest logistics facility in the Kansai region, with total floor space of roughly 110,000 square meters and inbound/outbound processing capacity of over 500,000 items/day each. At its core is the Amazon Robotics system: roughly 3,000 drive robots deployed throughout the warehouse, moving about 30,000 dedicated shelving units — the company says this allows for up to 40% more product storage compared to fixed shelving. Automated paper-bag packing equipment has also been introduced in the packing process. Amazon now operates more than 25 FCs and more than 65 delivery stations in Japan, with investment in Hyogo Prefecture alone exceeding JPY 44 billion in 2025. This is the most solid real-world deployment in this issue — not a pilot, but full-capacity operation live from day one.Source: BigGo Finance source
Waymo Removes Dallas Waitlist, Opens Citywide · autonomy
Anyone can now download the app and hail a ride, though rides to Dallas Love Field airport are not yet supported — that stretch of road is still under testing. Waymo opened public service in Dallas this February but with a waitlist requirement; during that period, roughly 150,000 riders used its app. The fleet consists of Jaguar I-Pace vehicles, with charging and maintenance handled by Avis Budget. Removing the waitlist follows the same playbook Waymo has reused in Phoenix, Los Angeles, and San Francisco. Worth noting is its Texas track record: this past May, service was temporarily suspended in Dallas, Houston, San Antonio, and Atlanta due to issues handling heavy rain and flooded roads.Source: TechCrunch source
Humanoid Robot Daily Rental Business Takes Shape: AgiBot's SHAREBOT Surpasses 5,500 Orders in Three Months · humanoid
AgiBot (Chinese humanoid robotics startup) established rental subsidiary SHAREBOT in December 2025; in its first three months live, the platform logged over 5,500 orders, with daily rental prices starting around RMB 3,500, and that price includes transport and a human operator on-site. One e-commerce livestreamer in Hangzhou rents out her own robot at RMB 3,000/day for trade shows, event performances, and even marriage proposals, having purchased the unit for roughly $30,000 in 2025; an investor in Jiangsu bought 8 units specifically to rent out. "What's actually being rented is a small service team, not an autonomous laborer" — the fact that the operator is included is itself telling: the current commercial loop runs on event-exposure value rather than autonomous work capability. AgiBot projects this market could reach $1.5 billion by the end of 2026, a company forecast.⚠️ Vendor's own accountSource: Bitcoin News source
Aurora Expands Driverless Freight into Refrigerated Transport · autonomy
Aurora has expanded its partnership with Charger Logistics to bring driverless truck capacity into refrigerated freight. Cold chain demands higher timeliness and temperature-control continuity than general freight, marking an extension of autonomous long-haul trucking into a higher-value category.Source: FleetOwner source
EACON Autonomous Driving System Installed on New-Energy Truck Fleet at Xinjiang Mining Site · autonomy
Easy Control Intelligent Driving's (Chinese mining autonomy company) autonomous driving system has been deployed on a new fleet of Lovol pure-electric and KNOW-HOW methanol trucks in Xinjiang. Mining sites are among the closed-scenario applications where autonomous driving has already achieved commercial viability in China, and the notable point here is that the system was installed simultaneously on two new-energy chassis types — pure-electric and methanol — rather than retrofitted onto conventional diesel trucks.Source: International Mining source
IV. Industry Developments
Nvidia Open-Sources Alpamayo 2 Super, 34B Autonomous Driving VLA Weights Ready for Direct Commercial Use · autonomy
The significance of this release isn't the parameter count but the license: the weights are released under the Linux Foundation's permissive OpenMDW-1.1 license, code under Apache 2.0, covering fine-tuning, derivative models, and commercial redistribution — and Nvidia is applying this license retroactively across the entire Alpamayo family, meaning versions previously available for research only can be deployed directly starting today. The model itself is a 34B vision-language-action model, combining a 32B VLM backbone built on Cosmos 3 Super Reasoner and post-trained with reinforcement learning, plus a 2.3B diffusion-based action decoder; a single forward pass takes surround-view camera video and simultaneously outputs a planned trajectory, a causal explanation for that trajectory, and meta-actions (yielding, lane changes, etc.). Training data comprises roughly 115,000 hours of multi-camera driving video and about 3.7 million Chain-of-Causation annotations. Nvidia reports a Lingo-Judge score of 79.2 on LingoQA, ranking first among nearly 40 evaluated models — 17.0 points above Qwen2.5-VL 72B and 15.1 points above Gemini 2.5 Pro; open-loop minADE₆@6.4s is 0.911 meters. The design explicitly targets long-tail scenarios: temporary rule changes at construction zones, vehicles cutting in illegally — situations that can't be exhaustively enumerated. Wu Xinzhou, Nvidia's head of autonomous driving, stated the strategic positioning explicitly for the first time — "Nvidia wants to be the Android of autonomous driving" — meaning it won't operate as a Robotaxi operator, but will provide the foundation of "three computers": in-vehicle inference, cloud training, and simulation verification; open-sourcing the model, the Cosmos world model, data tools, and evaluation framework together is precisely the move to lower the barrier to entry into this ecosystem. Both the leaderboard results and the claim of cutting "annotation cycles from months to days" are Nvidia's own self-reported figures.⚠️ Vendor's own accountSource: MarkTechPost, 36Kr source
Daegu, South Korea Completes the Country's First Humanoid Robot Safety Certification Center · humanoid
The city of Daegu announced on the 5th that starting this month it will fully launch the "Humanoid Robot Safety Certification Center" project, selected through a competitive process by the Ministry of Trade, Industry and Energy, which will establish for the first time in South Korea a safety verification and international standards certification system for humanoid robots. Over the five years through 2030, KRW 18.7 billion will be invested (including KRW 9.7 billion in national funding), located at the National Robot Testing Field in Dalseong-gun, led by the Daegu-headquartered Korea Institute for Robot Industry Advancement, with five other specialized institutions participating. The center will establish three evaluation systems — dynamic stability, AI reliability, and cybersecurity — linked with the National Robot Testing Field to provide one-stop support spanning test evaluation, demonstration, and certification, aiming to help Korean companies obtain safety certification and open overseas markets. Read alongside today's paper on the "fail-passive gap," this kind of national certification infrastructure is precisely the missing piece for humanoids entering industrial settings.Source: Seoul Economic Daily source
Samsung's Humanoid Push Revealed: Data Factory in Gumi, In-House World Model and Actuator Development in Parallel · humanoid
According to Korean media reports on the 5th, Samsung Electronics has confirmed its subsidiary Rainbow Robotics (Korean humanoid robotics maker, KAIST spinoff) is independently developing humanoid robots. Rainbow Robotics originated from KAIST's HUBO lab, with Samsung progressively acquiring a stake and gaining control in 2023–2024; its main revenue currently comes from the RB series collaborative robot arms, while it advances commercialization of the RBY1 wheeled bimanual humanoid platform. The newly disclosed development is Samsung's plan to build a data factory in Gumi, North Gyeongsang Province, integrating process data from existing production sites into a system to assist robot learning; earlier reports also indicated Samsung plans to build a humanoid mass-production base in Gumi, with total investment of roughly KRW 13 trillion. On the model side, industry sources say Samsung is also developing a world model; on the hardware side, Samsung is developing actuators in-house, drawing on its home-appliance motor technology. Organizationally, the RX (Robot Experience) business promotion office, established July 21, reports directly to CEO Roh Tae-moon, headed by Lee Dong-gun, who previously led Boston Dynamics' robotics strategy at Hyundai Motor Group.⚠️ Single-source accountSource: Cailianshe source
Mitsubishi Motors to Convert Idle Kyoto Engine Line, Targeting Up to 1,000 Humanoids per Month · humanoid
Mitsubishi Motors is partnering with Tokyo-based startup Highlanders to convert an idle engine production line at its Kyoto plant into a robot assembly line, targeting production start in early 2027 with a planned monthly capacity of up to 1,000 industrial humanoid robots. The first batch of robots will initially be deployed on Mitsubishi's own engine production lines to generate operational data, with external sales to follow validation. Highlanders is a University of Tokyo spinoff founded in 2023, whose flagship platform HL Human was released in 2025 with 19 degrees of freedom. Both capacity figures and timeline are per the companies' plans.⚠️ Planned figuresSource: Interesting Engineering source
Dobot Unveils Humanoid Robot Lumo, Targeting Home and Educational Companionship · humanoid
Dobot (Chinese robotics/robot-arm maker) unveiled the humanoid robot Dobot LUMO on August 5, standing nearly 1.3 meters tall and built on the company's in-house Kongyi embodied foundation model. Official demonstrations show it walking across outdoor terrain including lawns, sand, and cobblestones, along with emotion recognition, dancing, and backflips, while also being positioned as an embodied-AI teaching platform supporting further development. The company describes it as "the world's first fully all-terrain embodied humanoid robot," with the official demo video claiming no special effects were used — but "all-terrain" (quan qi) is a self-coined category term by the company, and its capability boundaries and mass-production timeline remain undisclosed.⚠️ Vendor's own accountSource: Securities Daily source
China Association of Automobile Manufacturers Establishes Autonomous Vehicle Industry Development Joint Council · autonomy
The China Association of Automobile Manufacturers has launched the establishment of an Autonomous Vehicle Industry Development Joint Council. Coming right after China's mandatory autonomous driving national standard just took effect (previously reported), this standing industry-association-level coordination mechanism typically corresponds to industry-wide alignment work on implementation details and compliance practices for the standard.Source: Sohu source
JD.com and Wuba Intelligence Reach Strategic Partnership · adjacent
The two parties will jointly advance the industrialization and application of embodied intelligence robots. E-commerce and logistics platforms continue to bring in embodied AI capability through strategic partnerships rather than in-house development — essentially locking in supply for their own warehousing and fulfillment scenarios.Source: Cnbeta source
Hardware · Supply Chain
· GigaDevice (Chinese chipmaker) GD32F50MxxG: a highly integrated motor-control MCU for humanoid joints, collaborative arms, and dexterous hands, packaging an in-house three-phase gate driver and 4-channel rail-to-rail op-amp into a single 252MHz Cortex-M33 chip, eliminating the need for external standalone drivers and sampling op-amps; already certified to IEC 61508 SIL2 functional safety, with the 8mm×8mm QFN80 variant now sampling and slated for mass production in December 2026, and the 7mm×7mm BGA100 variant planned for volume supply in March 2027 source
· Molex MiniMix: a hybrid connector combining 15.0A power delivery and 1Gbps automotive-grade Ethernet into a single interface, with a 5.65mm wiring profile that the company says takes up to 50% less routing space than alternative solutions, intended to replace the currently common manual wiring and hand-soldering inside humanoid joints — assembly labor time, not component cost, is the mass-production bottleneck, which is where this part's value lies (vendor's own account) source
· D-Robotics (Chinese embodied-AI chip maker) × GigaDevice: jointly released a full-stack control solution for a six-axis collaborative robot arm based on the Sunrise S600 processor plus GD32H77R MCU, with a Chinese compute chip and a Chinese control MCU now being supplied together as a complete package to robot-arm manufacturers source
· Hesai Technology (Chinese lidar maker): selected for Morgan Stanley's "Humanoid 100" list, with a lidar manufacturer now included in the humanoid supply-chain investment pool, reflecting the growing weight of its robotics business in its valuation narrative source
Top comments (0)