Today's Highlights
· Draft amendment to China's Road Traffic Safety Law adds a dedicated chapter on autonomous driving; violations to be handled by manufacturers and importers
· Skild AI lands $1.4 billion Series C at a valuation over $14 billion, led by SoftBank
· Gatik closes $200 million Series D, has completed 85,000 fully driverless deliveries
· Tiangong robot runs 100m in 8.86 seconds, half a second faster than its opening-ceremony record
· New FCC rule: robots assembled in the US must exceed 65% domestic component content by 2028
Research Progress
DELE-w0.5: World models don't need to render video first · world-model
Video generation is an unnecessary intermediate objective for world action models — this paper turns that claim into an experiment. DELE-w0.5 skips frame-by-frame prediction and infers action sequences directly from compressed future latent states, on the reasoning that what a robot needs is what the world looks like after the action completes, not what it looks like at every intermediate moment. Across four long-horizon manipulation tasks and 480 real-robot trials, overall task success reached 62.5, 47.5 percentage points above the strongest baseline, while cutting both training and inference cost.
Fenghao Lei et al. · arXiv 2608.22067 source
GeoWAM: A driving world model that predicts point clouds, not images · autonomy
Pixels entangle geometry, motion and lighting together, yet driving world models must infer 3D transformations from 2D images. GeoWAM replaces the state space so the pretraining target becomes the future scene's point-cloud geometry rather than future images, with an action head predicting the ego-vehicle trajectory conditioned on this geometric dynamics. Both open-loop and closed-loop evaluations show driving policies built on geometric world modeling clearly outperform image-based counterparts.
Yiren Lu et al. · arXiv 2608.23486 source
LD4WAM: Feeding 5,000 hours of human video into a world action model · world-model
Human video is cheap and diverse, but it can only teach pixel-level dynamics that aren't directly executable; action retargeting, in turn, leaves a large cross-embodiment visual gap. LD4WAM proposes motion-aligned latent dynamics as an intermediate representation — preserving full future video generation on one side, while using learnable queries to distill this dynamics layer to condition the action. The pretraining corpus is a self-built dataset of over 5,000 hours of human and robot data, tested successfully on the RoboTwin simulator as well as on real grippers and dexterous hands, holding up even on unseen objects and backgrounds.
Zhenhao Shen et al. · arXiv 2608.22403 source
Consistency loss placed in the wrong spot can hurt — and exposes inflated published numbers · world-model
Change the camera angle and world action models fall apart. This paper offers a near-counterintuitive finding: applying cross-view consistency loss to outputs that legitimately vary with viewpoint — like predicted future frames — is provably harmful; the correct approach is to constrain only the viewpoint-invariant parts, such as action chunks, future proprioception, and value. The method requires no camera calibration, extrinsics, or depth; on held-out viewpoints outside the training view envelope, closed-loop success is 12.2 percentage points above the paired control, with no gain within the envelope. A companion cross-backbone audit finds that previously published camera-robustness numbers were confounded by wrist-camera pose stability.
Bingqi Huang et al. · arXiv 2608.21402 source
The Imitator Game: Nine SOTA models collectively collapse at level four · benchmark
Humans imitate intent; robots learn trajectories. This benchmark splits that gap into four levels, L0 through L3, progressively widening the distance between human demonstrations and the robot's own scenario. Nine SOTA models perform steadily from L0 to L2, then collapse across the board at L3, where the bottleneck is functional substitution — using a different object's affordance to achieve the same intent. The accompanying IG-10K, currently the largest environment-aligned human-robot paired dataset, contains over 20,000 paired clips across 50-plus tasks and 6 domains, spanning both real robots and simulation; all models score a zero-shot success rate below 13% on unseen tasks, while fine-tuning on just 10 paired demonstrations yields significant gains that scale with pretraining size.
Xunzhe Zhou et al. · arXiv 2608.22301 source
Adding one retrieved sentence to the prompt drops VLA success rate from 92.47% to 3.00% · vla
The authors call this phenomenon prompt-form collapse: changing the form of an instruction alone is enough to dominate execution, and whether the appended semantics are meaningful is secondary — both meaningful text and length-matched nonsense text fail identically across all 500 states. TOWN-VLA's countermeasure adds an authorization gate to the prompt: only compact instructions meeting fixed compatibility rules are allowed to rewrite the policy input, otherwise the original prompt is restored byte-for-byte. On LIBERO-Plus, with 10,030 episodes per method, success rose from 69.5% to 73.1%; on a real PiPER robotic arm paired with a frozen π0.5 checkpoint, success rose from 52.7% to 78.7% across 150 trials.
Zhiruo Zhou et al. · arXiv 2608.23224 source
GOLEM: A humanoid disassembles retired EV battery packs, open source, with a capability ladder · manipulation
Disassembling scrapped battery packs is tedious and dangerous work, currently done almost entirely by hand. GOLEM decomposes walking, manipulation, dynamic balance, navigation, and spatial memory on a Unitree H1-2 humanoid into independent, interface-abstracted modules, wrapped in a Dockerized ROS 2 setup, with MuJoCo and IsaacLab digital twins matching the real robot's interface, all open source. The real value is its step-by-step evaluation ladder: grasping already-loosened fasteners from a real Hyundai Ioniq 5 battery pack, success reaches 97% while suspended, drops to 87% when freestanding, and falls to just 37% once pose deviation from navigation is layered on.
Max Conway et al. · arXiv 2608.21550 source
A single photo of a clinic room generates a simulated scene robots can be tested in · benchmark
Embodied AI needs to be tested in the clinical environments it will actually work in, but building a realistic, robot-ready site is expensive and hard to scale. Researchers validated the approach on 39 ophthalmology clinic scenes, converting a single photo into an editable environment ready for direct use in a simulator, then adding equipment meshes, collision proxies, and semantic anchors to turn visual reconstruction into a scene where contact can be computed. Three robot morphologies show clearly different reachability and contact feasibility on the same set of task goals; small translations and rotations in equipment position change the contact margin in ways that visual similarity alone cannot reveal.
Xinyuan Wu et al. · arXiv 2608.21416 source
InstructMove: A manipulation benchmark that fails you if you don't read the instruction · benchmark
Most manipulation benchmarks can't tell whether a policy is actually listening to instructions, since the target object is often the most visually salient one, or physically only one action is feasible — so guessing still works. InstructMove requires multiple actions in the scene to be both visually and physically feasible, with only one consistent with the language, then splits instruction-following into four categories: category recognition, attribute discrimination, spatial reasoning, and compositional pick-and-place. Experiments show it can diagnose visual shortcuts a policy is taking, and its simulated data also improves real-robot instruction-following performance; the code is hosted in the Horizon Robotics RoboOrchardSim repository.
Mengao Zhao et al. · arXiv 2608.22990 source
Other papers today: WorldToken fuses multi-view images, proprioception and task conditioning into a single world token at each timestep; an 85.3M-parameter policy reaches 59.45% closed-loop success across 23 RoboCasa tasks (arXiv 2608.22591 source); WAM-OPD uses on-policy distillation to repair world action model students whose capability degraded after acceleration (arXiv 2608.22364 source); DreamMimic uses a world model to teach humanoid whole-body visual loco-manipulation (arXiv 2608.22278 source); DELTA uses a deformable elevation attention encoder to decode sparse terrain foot placement for quadrupeds (arXiv 2608.22033 source); CounterAlign adds counterfactual negative supervision to VLA behavior cloning that otherwise only has positive samples (arXiv 2608.21740 source); GuardianBench fixes the scene and varies only the instruction to test embodied models' latent situational risk (arXiv 2608.21928 source); MomADv2 suppresses inconsistent temporal memory when driving instructions change (arXiv 2608.23405 source); OpenSCvx is an open-source Python framework for nonlinear trajectory optimization (arXiv 2608.21631 source); Astrobee grayscale images plus 3D Gaussian Splatting reconstruct the ISS's Japanese experiment module (arXiv 2608.21685 source).
Open Source · Tools · Benchmarks
· RoboColiseum: a standardized embodied-AI simulation evaluation platform launched by a Shanghai team, with task suites and leaderboards across four capability dimensions; the team says the sim-to-real gap for the same model is under 10%, and hundreds of teams have already run models on it during closed testing source
· HOMIE Gen2: a head-mounted capture rig from Singapore's Ropedia, with four cameras, spatial audio and inertial data hardware-synchronized to within 50 microseconds; recordings include camera pose, depth, hand and body keypoints, and task annotations — the company says it is the first first-person robot training capture device with a 360-degree field of view source
Funding & Deals
Skild AI ｜ Series C ｜ $1.4 billion ｜ post-money valuation over $14 billion · embodied
Led by SoftBank, with participation from Nvidia, Bezos Expeditions (Jeff Bezos's investment vehicle), Macquarie Capital, and 1789 Capital, plus Samsung, LG and Schneider Electric joining as strategic shareholders. The company, three years old, has raised over $2 billion total. Skild doesn't build hardware — it sells Skild Brain, software that drives quadrupeds, humanoids, desktop robotic arms, and mobile manipulation robots; ABB, Universal Robots, and Mobile Industrial Robots have already embedded it in their machines. The software currently runs on hundreds of robots across factories, data centers, and logistics hubs, including deployments at an Nvidia factory in Houston and a pilot at LaGuardia Airport. In under half of 2025, Skild recorded roughly $30 million in revenue. Co-founder and president Abhinav Gupta, who left Meta's FAIR robotics lab to found the company with Carnegie Mellon professor Deepak Pathak, said "this is the beginning of the GPT moment, and it's going to happen over the next year to a year and a half" — the bet is on factories and warehouses, not homes.Source: Dealroom source
Gatik ｜ Series D ｜ $200 million · autonomy
Co-led by the Qatar Investment Authority and Koch Disruptive Technologies, with participation from Millennium Management, ARK Invest, and Intact Private Capital. Gatik runs short- and mid-haul driverless freight between distribution centers and stores, with routes across Texas, Arkansas, Arizona and Canada, serving customers including Walmart, Tyson Foods, and Canadian grocery chain Loblaws. The company says it has completed 85,000 fully driverless deliveries, with over $600 million in contracted revenue and a 99% on-time rate. In January, CEO Gautam Narang told Forbes the fleet would grow from just over a dozen vans to hundreds of vans by the end of 2026.Source: Forbes source
Generalist AI ｜ new round ｜ $200 million · embodied
Led by 8VC, with participation from some existing investors; valuation undisclosed, coming just two months after its $400 million round in June. About a week earlier, the company released Gen-1.5, aimed at compressing the setup time for factory automation workflows: a user wearing sensors or using the robot's own camera demonstrates a task once, and the model learns the new task — across 10 sample tasks, average completion rate was 59% from a single demonstration (previously reported).Source: Axios (via Digital Today) source
Airbound ｜ Series A ｜ $37 million · adjacent
Led by Greenoaks, with participation from DoorDash, Lightspeed, Lachy Groom, and Humba Ventures, less than a year after an $8.65 million seed round, bringing total funding to nearly $50 million. This Indian company builds tail-sitter VTOL drones designed so the aircraft weighs less than its payload: the current model, TRT, weighs about 3.3 pounds and carries 2.2 pounds, while a model in development weighs about 6.6 pounds and carries up to 11 pounds. Over 13,000 autonomous flights have been completed in Bangalore and Guntur, including more than a thousand delivering lab samples for the Narayana Health hospital network — a 2.5-mile trip taking about 7 minutes, versus 3 to 5 hours by two-wheeler on the ground including batching and wait time.Source: TechCrunch source
General Intuition ｜ new round ｜ valuation around $6 billion · world-model
Following reports on the 23rd that talks were underway, the investors have been revealed as Valor and Point72, as the company simultaneously shifts its focus toward robotics.Source: TechCrunch source
NEC × Dexmate ｜ strategic investment ｜ amount undisclosed · embodied
NEC's Orchestrating Future Fund has invested in US robotics company Dexmate, maker of the VEGA humanoid robot.Source: NEC source
Commercialization & Deployment
Galbot S1 (Chinese robotics startup) runs 7x24 on a CATL production line · embodied
Galbot S1, with a maximum dual-arm payload of 50 kg, is now running 7x24 on a CATL production line, handling depalletizing, transport, and palletizing of materials. At the same WRC event, Galbot also unveiled its humanoid robot Xingzai and the embodied foundation model that drives it, AstraBrain, with the company saying the same general-purpose brain can control bipedal, wheeled, and heavy-load robot bodies simultaneously. A booth demo and stable running on a production line are two different things — the former's reliability cannot be directly extrapolated to the latter.⚠️ Vendor claimSource: Gasgoo source
JAKA Robotics' (Chinese robotics manufacturer) embodied-AI revenue hits 45 million yuan in H1, up more than 20-fold year over year · industrial
JAKA Robotics reported H1 revenue of about 320 million yuan, up 106.6% year over year, with embodied-AI business revenue topping 45 million yuan, up more than 20-fold year over year and accounting for nearly 15% of total revenue; the company has 231 embodied-AI customers to date. R&D spending was about 100 million yuan, up 148.4% year over year, widening overall losses accordingly. The company says its self-developed DobotWAM model achieved a 99.25% average success rate on LIBERO, higher than mainstream published models as of results available through 2026.⚠️ Vendor claimSource: Sohu source
Keenon Robotics' KOM 3.0 packs a latent-space world model into service robots · embodied
The prior generation, KOM 2.0, used a fast-slow dual system: the slow system, K-Mind, handled task understanding and planning, while the fast system, K-Act, handled fine-grained motion generation. KOM 3.0 builds on this by integrating a latent-space world model, which Keenon calls the world's first VLA architecture for the service industry to incorporate a latent-space world model. At the WRC venue, the company set up four scenarios — a coffee kiosk, a dessert shop, a retail store, and a laundromat — where its XMAN-R1 humanoid, acting as a guest barista, autonomously completed the full workflow from cup retrieval to extraction, while two humanoids in the laundromat coordinated washing and folding under unified scheduling. COO Wan Bin said cumulative shipments have exceeded 100,000 units, with business covering more than 60 countries and regions.⚠️ Vendor claimSource: Gasgoo source
Amazon planning fully automated delivery stations "Tetromino"; company says document figures are inaccurate · industrial
An internal planning document from last month, seen by Business Insider, states that Amazon plans to invest $103 million in its first Tetromino pilot in 2028, build five sites in 2029 at about $85 million each, add ten more in 2030, and reach cumulative investment of over $530 million by 2029. The document says the design's package-processing rate is about 2.5 times that of existing delivery stations. An Amazon spokesperson told BI this is an "early-stage concept" and that the specific financial figures and roadmap in the document are "inaccurate and do not reflect our current plans." The delivery station is the last step before packages reach drivers, and put-away and loading are widely regarded in the industry as the hardest and most expensive parts to automate. The document mentions that one key technology may come from startup Boxbot, whose pallet storage-and-retrieval system claims to speed up loading by up to tenfold. On its July earnings call, Amazon said the number of robotic arms would more than double in 2026.Source: Business Insider source
Waymo selects Munich as its first EU robotaxi city · autonomy
The Financial Times reports Waymo has designated Munich as its first EU launch city, having already registered a German subsidiary, Waymo Germany GmbH, in June. Separate reports say it has also set up a presence in Madrid to prepare for autonomous-taxi testing. On the same day, Amazon's Zoox expanded its steering-wheel-free robotaxi from Las Vegas to additional cities including San Francisco (previously reported as having started charging fares).Source: Financial Times source; Olive Press source; NBC News source
UBTECH and Singtel sign memorandum of understanding · humanoid
The two companies signed an MOU at the 2026 World Robot Conference aimed at combining embodied-AI robots with telecom service capabilities, to advance robot deployment in Asia-Pacific enterprise operations. The memorandum did not include a dollar amount, delivery volume, or timeline.⚠️ Planning stageSource: Beijing Business Today source
Industry Developments
Draft amendment to China's Road Traffic Safety Law adds dedicated autonomous-driving chapter; violations handled by manufacturers and importers · autonomy
The draft amendment to the Road Traffic Safety Law was submitted on the 25th for its first reading at the 24th session of the Standing Committee of the 14th National People's Congress, running 9 chapters and 170 articles in total. The autonomous-driving section contains four articles: self-driving vehicles must pass a road-traffic-rule compliance test and be registered in accordance with the law before they can operate on public roads; traffic violations occurring while the autonomous driving system is operating are to be handled by the manufacturer or importer, with specific procedures to be drafted by China's public security authorities under the State Council; a compulsory third-party motor vehicle traffic accident liability insurance system will apply to self-driving vehicles, with rules to be set by the State Council; and manufacturers and importers must ensure road safety, cybersecurity and data security, and may not falsely or exaggeratedly advertise autonomous-driving functions. The current Road Traffic Safety Law took effect in 2004, and its three amendments in 2007, 2011 and 2021 each touched only isolated provisions. Shifting the party responsible for violations from the driver to the manufacturer or importer would directly rewrite the compliance cost structure for autonomous-driving companies operating in China. The draft also includes provisions capping e-bike speed at 20 km/h, banning "distracted riding," and regulating so-called "roving mob" group cycling.Source: China News Service source
FCC adds advanced mobile robots to Covered List; US-made robots must exceed 65% domestic content by 2028 · hardware
The US Federal Communications Commission added advanced mobile robots to its Covered List in July. Per the FCC website's Q&A page, new advanced robots assembled in the US must have domestic component value exceeding 65% by 2028, rising above 75% by 2029. Regulators have yet to clarify how exemptions will work. Jeff Burnstein, president of A3, an industry association representing nearly 1,500 robotics and automation companies, said "everyone is trying to figure out what this means for them and what the impact will be," calling the regulatory environment "a day-to-day thing right now." Peter Cowhey, professor emeritus at UC San Diego and a former FCC official, told Nikkei that the rule's intent is to diversify the supply chain and bring some capacity back onshore, but "the problem is there's no concrete mechanism to get from here to there, to swap Chinese products for American ones" — calling it "an incoherent policy of forced import substitution."Source: Nikkei Asia source
Tiangong robot clocks 8.86 seconds in the 100m · humanoid
In the first heat of the men's large-class 100m at the second World Humanoid Robot Games, Tiangong, built by the Beijing Innovation Center for Humanoid Robotics, ran 8.86 seconds, shaving over half a second off the 9.39-second record it had just set at the opening ceremony. Reuters covered the same event under the headline that a Chinese robot ran under 9 seconds in Beijing. Elsewhere at the games, things were less dignified: a humanoid collided with the judges' table during a weightlifting event and was carried off on a stretcher.Source: Sichuan Observer source; Reuters source; EssentiallySports source
XPeng's second-generation VLA upgraded to 6.3.0, on-device parameter count up 3.5x · autonomy
The new version increases the on-device model's parameter count to 3.5 times the previous version and improves end-to-end response speed by 300%; XPeng says this scale is 15 times that of mainstream VLA models in the industry. The launch vehicle, the G9L, uses XPeng's Turing AI chip — one chip for the Max trim, two for Ultra SE and Ultra. Version 6.3.0 brings some robotaxi-grade L4 experiences into production vehicles for the first time, including smart standing starts, precise curbside stops, and autonomous parking-spot search in campus settings; the system supports over 30 seconds of temporal reasoning and can anticipate scenarios up to 6 seconds ahead. The VLM has also been added to the cabin, supporting vague-destination navigation and voice-guided nearby parking. The G9L has accumulated over 6.74 million km of R&D testing mileage across 26 countries and regions, with rollout starting in late August.⚠️ Vendor claimSource: Netcarshow source
He Xiaopeng: IRON to enter mass production by year-end, formal launch in 2027 · humanoid
At the Q2 earnings call, XPeng CEO He Xiaopeng gave a timeline: mass production begins by the end of this year, starting with rollout at XPeng's own stores and campuses; the IRON humanoid formally launches in 2027, aiming for large-scale delivery to external retail and service-industry customers in China and overseas, with monthly production capacity scalable to several thousand units next year depending on market demand. He said IRON has 76 degrees of freedom across its full body, with 21 in each dexterous hand, and that the robotics R&D effort has spanned eight years, with the technical challenge of a high-end general-purpose humanoid at least 20 times greater than that of a smart car. He also predicted that once IRON launches, per-unit lifetime revenue and gross-profit contribution will be far higher than the current per-vehicle average price and margin in XPeng's auto business. The humanoid unit completed its first funding round last week, raising over $900 million at a post-money valuation above $6.3 billion (previously reported).⚠️ Executive remarksSource: Kuaikeji (via Sina Tech) source
NIO's autonomous-driving chief Ren Shaoqing founds a physical-AI foundation model company · world-model
At an autonomous-driving all-staff meeting on August 24, NIO CEO William Li announced that Ren Shaoqing has founded an independent company focused on physical-AI foundation models and embodied intelligence; NIO will make a strategic investment and collaborate with the new company, while Ren continues to head NIO's autonomous-driving business. The new company has been registered, though its name and funding details have not been disclosed. Ren is one of the four authors of ResNet and one of the creators of Faster R-CNN; after joining NIO he led the mass-production delivery of NWM, described as China's first large-scale, forward-simulating world model for autonomous driving. 36Kr separately reported the new company achieved unicorn status upon registration, which does not square with reports that its funding remains undisclosed.Source: LatePost (via ThePaper/Zhidongxi) source; 36Kr source
South Korea launches K-Moonshot AI humanoid initiative, aims to name a national-team platform within three years · humanoid
South Korea's Ministry of Science and ICT held a meeting on the 25th to gather input from industry, academia and research institutes on a draft "K-Moonshot AI Humanoid Roadmap," with the goal of producing a world-class national-team K-humanoid platform within three years, localizing AI models and key components, and running the effort from core technology development through demonstration and mass production. Also announced was the Next-X Humanoid Challenge, which will evaluate humanoid robots developed in South Korea and select the best performers for support, with the challenge ultimately choosing a national-team-level platform starting in 2027. The initiative's overall lead, Yeo Jun-gu, said that after AI, the next wave is physical intelligence — systems that can judge and move autonomously in the physical world — with humanoids as its core platform. The roadmap will incorporate feedback from this session before going to a meeting of science-related cabinet ministers, with a final version due next month.Source: Digital Today source
Shenzhen industrial robot output up 45.5% year over year, January–July · industrial
According to Shenzhen Municipal Bureau of Statistics data released on the 24th, industrial value-added output above designated size in the city rose 9.2% year over year for January–July, 0.5 percentage points faster than the first half of the year, with specialized equipment manufacturing up 14.4% and general equipment manufacturing up 14.3%. Among high-tech products, output of 3D printing equipment, industrial robots, and lithium-ion battery products rose 62.3%, 45.5%, and 27.1% respectively.Source: Cailianshe (via Sina Finance) source
Hardware · Supply Chain
· Actuators: A report from Gavekal Technologies dated August 10 says actuators typically account for 40%–60% of a humanoid robot's hardware BOM, with batteries, lidar and motor controllers also mostly sourced from China — cited as the direct reason the US domestic-content threshold will be hard to meet (via Nikkei Asia) source
· Dexterous hands: Forbes visited Beijing-based Linkerbot (Chinese dexterous-hand startup), which the company says was valued at $3 billion in an April Series B+ round backed by Ant Group, HSG, and CICC Capital; a report from GGII estimates Chinese dexterous-hand shipments will reach about 70,000 units this year, up nearly 300% from 2025, expanding to 430,000 units annually by 2030; Tianfeng Securities estimates over twenty Chinese companies were already making robotic hands as of 2025 source
· AAEON robot development platform: a robot development platform based on Intel Core processors has launched, targeting edge-side compute configurations source
