Today's Highlights

· World Labs releases world model Atlas, generating a controllable one-minute 1440p video plus 3D output from a single image

· Uber cuts 3,300 jobs, freeing up funds for over $10 billion in autonomous driving partnerships

· Perception company Lyte raises a $165 million Series C at a $1.6 billion post-money valuation

· Unitree's (Chinese quadruped/humanoid robot maker) stock has halved from its first-day price on its 11th trading day since listing, market cap RMB 220.85 billion

· The world's first international standard for legged robots, led by Chinese experts, is officially released

· The embodied data gap exceeds 99%; 25 Chinese data-collection companies raised over RMB 17 billion in H1

Qwen-Drive-1.0: Packing Driving Ability into a Vision-Language Foundation Model Without Losing General Capability · autonomy

HuggingFace popularity 297, the highest among today's papers. End-to-end driving approaches typically trade general vision-language capability for driving capability. Qwen-Drive-1.0 leaves the architecture untouched and instead attaches a BEV perception head that simultaneously handles 3D detection, semantic occupancy prediction, and map segmentation, with a Planning Expert generating ego-vehicle trajectories from the shared representation. Staged training mixes driving supervision with general vision-language data; the authors report competitive motion-planning performance across open-loop, pseudo-closed-loop, and closed-loop settings, while largely preserving general vision-language capability.

Xin Zhou et al. · arXiv 2609.00111 source

ZimaBlue: Feeding in 120,000 Hours of Human First-Person Video Lifts Real-Robot Zero-Shot Success from 36.1% to 77.8% · world-model

Action-labeled robot trajectories are expensive and monotonous, while first-person video is cheap and abundant — but lacks action labels. ZimaBlue converts video into control ability through a three-stage curriculum: causal pretraining on human and robot first-person video, followed by video-action mid-training on heterogeneous robot trajectories using a unified action representation, then specialization for the target embodiment. In real-robot zero-shot evaluation, expanding training data from only the target robot's own data to 120,000 hours of video raised the success rate from 36.1% to 77.8%; an asynchronous Slow-Fast dual-system design lets the lightweight branch run at 30 Hz on an RTX 4090. HF popularity 38.

Xionghao Wu et al. · arXiv 2609.00188 source

IMPACT: The Supervision Signal in World Models Gets Swallowed by Static Backgrounds · world-model

Physically plausible interaction has long been a weak point for world models; prior fixes attached motion, geometric, or semantic representations, but these dense representations require extra estimators or manual annotation, capping training scale. This paper returns to the training objective itself, pointing out a supervision mismatch in the globally-averaged MSE denoising objective: static content, which dominates the frame, dominates the gradient, while the sparse dynamic regions that actually determine interaction quality are undersupervised. IMPACT treats cross-attention on the manipulated object's tokens as an internal spatiotemporal prior, sampling candidate regions and calibrating them with local prediction error into an interaction map, then reweights the denoising supervision accordingly — without external representations or changes to the inference pipeline. It consistently outperforms the corresponding MSE baseline across arm and hand manipulation and multiple DiT backbones.

Rongze Tang et al. · arXiv 2609.00161 source

REFACTOR-VLA: Building a Reusable Action-Skill Library for VLAs — Where a Bigger Model Performs Worse · vla

VLAs like OpenVLA, π0, RT-2, and RDT-1B are monolithic, outputting raw actions or short action chunks directly, and degrade on long-horizon tasks while remaining hard to interpret. This paper builds a wake/sleep system: during sleep, a learned implicit world model runs rollouts, from which behavioral equivalence is used to cluster action segments; during wake, typed lambda terms are issued and handed to a library-conditioned rectified-flow decoder for execution, with abstractions admitted to the library only after passing two gates — minimum description length and return preservation. Two findings run counter to intuition: scaling the world model from 188M to 430M parameters made performance worse across all four LIBERO suites; while adding an InfoNCE auxiliary contrastive loss for world-model warmup pushed clustering NMI an average of 0.184 above the strongest published baseline across the four suites.

Riyaaz Shaik, Chandru Venkataraman et al. · arXiv 2609.01215 source

DroneCATS: Small Models Reach the Target but Don't Know They've Arrived · benchmark

Multimodal large models are dropped directly into a drone control loop, with the action space declared entirely via prompt — no fine-tuning, no function calling. Testing four capabilities (approaching a visible target, tracking a moving target, turning to search for an out-of-view target, and commanding multi-drone formations) reveals a striking contrast: flying isn't the problem. Open-source models as small as 2B parameters are often more reliable at entering the success radius than frontier models, yet lose out by declaring arrival prematurely or not declaring it at all; in multi-drone command, small models blindly copy the same coordinate across different viewpoints. Spatial awareness passes; the action protocol doesn't. HF popularity 19.

Jaewoo Park et al. · arXiv 2609.01404 source

NegGrasp: Tell the Robot "Don't Grasp Here" and It Grasps There Anyway · manipulation

Language-driven dexterous grasping models like DextER perform well when the instruction specifies where to grasp, but fail systematically once a negative constraint is added, such as "grasp the handle but avoid the cup body." The cause lies in the training corpus: datasets like DexGYSNet contain almost no avoidance-type instructions, so the model learns that any mentioned part is a contact target, turning a forbidden zone into just another grasp point. This paper adds no negative training samples at all; instead, at inference time it uses sequential Monte Carlo with classifier-free guidance to steer sampling toward the specified part while pruning candidates heading toward the forbidden zone, with the forbidden zone's location read from language by a frozen 3D part-grounding model. On the accompanying NegGrasp benchmark, the violation rate of the strongest baseline drops from 57.9% to 17.2%.

Geonho Kim et al. · arXiv 2609.00555 source

CoLT-Drive: The Long-Tail Problem Isn't "Did It Recognize the Object" — It's "Did the Object Change What I Can Do" · autonomy

Long-tail failures in autonomous driving are typically framed as rare-object recognition errors; this paper argues the decision-level question is whether the model can infer how an object changes the ego vehicle's feasible high-level actions. To make small VLMs deployable, the authors propose the KPA framework, combining structured perception-to-decision prompting, SLERP-based expert merging, and a driving-condition-routed RegMoE LoRA mixture-of-experts module, aiming to adapt to long-tail cases without washing out the open-world knowledge in the pretrained model.

Zhengxu Tang et al. · arXiv 2609.00242 source

Lyte ｜ Series C ｜ $165 million ｜ $1.6 billion post-money · hardware

Led by Maverick Silicon, with Managing Partner Andrew Homan joining the board. The company works on the foundation layer of robot perception: proprietary sensor chips, multimodal sensors, and spatial software that compress 4D coherent vision, high-resolution imaging, and inertial sensing into a single synchronized timeline, measuring position and motion directly rather than reconstructing them in software. The founding team comes from Apple and PrimeSense, having worked on the 3D sensing behind Kinect and Face ID. CEO Alexander Shpunt's framing is that "physical AI has a perception problem before it has a model problem." The company came out of stealth only this January and has since entered mass production, shipping to robotics customers in inspection, logistics, and manufacturing. Source: Business Wire source

Locus Robotics ｜ Series G ｜ $41.6 million · industrial

Following yesterday's report that the round was nearing close, the amount is confirmed: $41.6 million, entirely from existing shareholders. CEO Rick Faulk said the round remains open and the company is still seeking additional funds from existing investors. This warehouse AMR company's last major raise was a $117 million Series F in November 2022, led by Goldman Sachs Asset Management and G2 Venture Partners, at a valuation near $2 billion. Ahead of this round, the company launched Locus Array and acquired Nexera Robotics. Source: FinSMEs source

Wandercraft ｜ In talks ｜ approx. €100 million ｜ approx. €750 million pre-money · humanoid ⚠️ unconfirmed report

Per Bloomberg, citing people familiar with the matter, with Goldman Sachs serving as financial advisor. This Paris-based company started in 2012 with rehabilitation exoskeletons; its Atalante X is already in use at over 100 hospitals and rehabilitation centers worldwide. Its new Calvin line pivots to industrial humanoids — the first model, Calvin-40, went from design to build in 40 days, runs on NVIDIA Isaac GR00T N1 and Jetson, and can repeatedly lift 40 kilograms. Renault is both a shareholder and the first commercial customer, already using Calvin-40 to mount tires on the production line at its Douai plant in France; the roadmap calls for about 10 units by the end of 2026 and 350 units by the end of 2027 across plants in France and Spain, targeting a 30% reduction in per-vehicle assembly time. Per Dealroom, humanoid robot companies had already raised $8.6 billion in H1 this year — 1.8 times the total for all of 2025. Source: Tech Funding News source

United Imaging Vision ｜ New round ｜ hundreds of millions of RMB · hardware

Investors included Dunhong Asset Management, the Hangzhou Embodied Intelligence Industry Fund, Fuyang Industrial Investment, and Zhipingfang (Chinese embodied-AI startup). This is a vision-solutions provider that has spent over a decade in the field, with customers including Unitree, Zhipingfang, AgiBot, Rokid, and Pico, while also supplying Ecovacs, Yunjing, and Trifo's robot vacuums and mowers — revenue is already at the RMB 2 billion level. Founder Zhang Feng started in phone camera modules in 2014, moved into robot vacuums in 2018, and pivoted to on-device AI vision in 2021. Funding will go toward its depth-camera product line, on-device visual understanding, data collection and evaluation capability, and a billion-RMB-scale smart vision industrial base in Hangzhou. Source: PEdaily source

Jaka Zhi Technology ｜ Hong Kong IPO ｜ updated prospectus · industrial

Revenue for H1 2026 was RMB 169 million, gross profit RMB 49.83 million, gross margin 29.4%, and a period loss of RMB 63.7 million. Over the prior three years, revenue was RMB 74.95 million, RMB 115 million, and RMB 266 million respectively, with period losses of RMB 114 million, RMB 118 million, and RMB 99.73 million — losses are narrowing but the company has yet to turn profitable. By revenue mix, standard AMR and omnidirectional heavy-load AMR solutions account for 54.7%, forklift-type AMRs 27.9%, and robot products only 10.2%. Board chair Xiong Rong controls 29.61% of voting rights, Zhongkong Technology founder Chu Jian holds 3.67%, and Legend Capital (Lenovo's investment arm) holds 5.80%. Cash and equivalents stood at RMB 218 million as of June 30. Source: 36Kr source

MirrorSpace ｜ Seed+ round ｜ over RMB 10 million · hardware ⚠️ compiled from public filings

Investors included Haiyuan Capital, Miracle Plus (formerly Y Combinator China), Songhe Capital, and Jinyili, with the information compiled from Qichacha public filings. This Dongguan-based company makes plug-and-play spatial perception modules, fusing multi-source inputs such as RGB, LiDAR, and thermal imaging into a dynamic 4D spatial representation. Source: Stockstar source

Hyundai Mobis Wins All Actuator Orders for the Atlas Humanoid, 31 Units per Robot · hardware ⚠️ analyst estimate

Actuators, which convert energy into physical motion, are the most expensive and difficult component in a humanoid robot, and Hyundai Mobis has secured the entire actuator scope for Atlas, with 31 body actuators per unit. LS Securities named Hyundai Mobis its top pick in South Korea's physical AI value chain that day, maintaining a buy rating with a target price of KRW 750,000, and estimated that at an annual production target of 30,000 units, the robotics business could generate revenue of KRW 1.04–1.43 trillion. These revenue figures come entirely from the brokerage's model, not company disclosures; Mobis's own robot-component revenue isn't expected to become significant until 2027. Source: CNMO source

DEEP Robotics' (Chinese quadruped/wheeled robot maker) Wheeled-Legged Robot Enters Turpan Vineyards, Reportedly Cutting Manual Labor by 70% · industrial ⚠️ manufacturer's claim

The machine doesn't pick grapes. DEEP Robotics' Lynx M20S instead carries baskets of already-picked grapes from between the trellis rows to a collection point, while also repositioning drip irrigation tubing and relaying field data. It carries a 35-kilogram payload and operates across a temperature range from 50°C to minus 30°C — ground temperatures during Turpan's August harvest season routinely exceed 50°C — and can navigate a 50-centimeter-wide aisle without turning around. Grapes begin losing moisture as soon as they're picked, so shortening the time from harvest to transport directly affects sellable quality, which is the practical reason the vineyard adopted the robot. The report cites "over 70% reduction in manual labor" but does not specify per-unit cost or the scale of rollout beyond Turpan. Source: Global Agriculture source

China National Heavy Duty Truck Group Delivers 30 Huanghe Smart-Driving Tractor Trucks, Signs Strategic Agreement for 200 More · autonomy

Delivered on August 31 in Huai'an, Jiangsu, to customer Leijie Logistics, which specializes in medium- and long-haul trunk routes within Jiangsu. The trucks feature L2+ level driver-assistance, not driverless operation; the selling points lean more toward cost-reduction specs such as an integrated tractor-trailer drag coefficient of 0.349 and a third-generation MC13 engine rated at 620 horsepower. China National Heavy Duty Truck Group sold 188,500 heavy trucks in H1 this year, up 26.69% year-on-year, for a 28.5% market share. Smart-driving orders in trunk logistics currently remain mostly assistance-level, on a different track from the L4 fleets used in mining and port operations. Source: Fangde.com source

Didi's Robotaxi R2 Comes Standard with 10 RoboSense LiDAR Units · autonomy

Following the R2's launch of fully driverless paid ride-hailing trials in Beijing and Guangzhou on August 31, supply-chain details have emerged: each vehicle comes standard with 10 RoboSense (Chinese LiDAR maker) LiDAR units. For upstream suppliers, per-vehicle LiDAR usage is an order of magnitude higher than in factory-installed passenger cars, making this one of the segments most directly benefiting from Robotaxi scale-up. Source: Gasgoo source

Waymo to Begin Robotaxi Testing in Greater Cincinnati · autonomy

Following the September 1 same-day launch of paid driverless rides in Denver, San Diego, and Tampa, Waymo is extending its testing footprint further into the Midwest. Source: The Business Journals source

World Labs Releases Atlas, Generating a Controllable 3D World from a Single Photo · world-model

Fei-Fei Li's World Labs has released its next-generation world model, Atlas, built on a multimodal autoregressive diffusion transformer architecture that treats text, images, camera poses, and 3D depth maps all as native inputs, with each input image anchored to a 3D position in space to jointly form a "spatial context." It can do four things: generate images and video with pixel-precise camera control, up to one minute at 1440p; reconstruct real 3D scenes from one to dozens of images, outputting point clouds or 3D Gaussian splats; jointly model space and time from input video; and generate images and 360-degree panoramas from text. In one example, feeding the model just 25 ground-level photos of Stanford's Main Quad, it generated a flyover path above the campus on its own. For robotics, Atlas offers a route to mass-producible real-to-sim conversion: the team used a phone to shoot two large environments, took 24 frames from each for reconstruction, then let simulated robots of different configurations move through the reconstructed scenes while Atlas generated, in real time, the RGB and depth data the onboard camera should see; after a task run, objects, positions, and lighting can be varied for data augmentation. Three synchronized cameras can also produce "bullet time"-style multi-view freeze frames. The team says Atlas follows a scaling law, with performance improving steadily with training compute, and that it will drive iteration on products like Marble. World Labs has raised $1.2 billion to date, with investors including NVIDIA, AMD, and Autodesk. Source: Zhidongxi source

Uber Cuts 3,300 Jobs, Redirecting Funds to Over $10 Billion in Autonomous Driving Investment · autonomy

The cuts amount to about 10% of Uber's global workforce, the largest since the 2020 pandemic layoffs. Management headcount was cut by about 20%, with some managers moving to individual-contributor roles, and the number of teams with only one or two people was nearly halved. CEO Dara Khosrowshahi's stated reason is that five years of rapid growth had built up excess layers, excessive coordination overhead, and diffuse accountability. The dining, retail, and direct-delivery operations groups have been merged, and engineering and science teams are also being consolidated. The company expects to invest over $10 billion in autonomous driving partnerships over the coming years; the Rivian partnership announced in March is ultimately meant to support up to 50,000 fully driverless robotaxis, launching first in San Francisco and Miami in 2028. Uber's stock rose more than 2% in pre-market trading following the announcement. Source: Analytics Insight source

Embodied Data Gap Exceeds 99%; 25 Data-Collection Companies Split Over RMB 17 Billion in H1 · adjacent ⚠️ compiled report

China Academy of Information and Communications Technology's "Embodied Intelligence Training Ground Research Report (2026)" assesses that globally available high-quality real-world data amounts to only hundreds of thousands to about a million hours, while embodied foundation models would need at least tens of millions of hours for capability emergence — a gap of over 99%. According to IT Juzi, 25 Chinese embodied-AI data startups raised a combined total of over RMB 17 billion in H1 2026. MiFeng Technology (spun out of AgiBot, Chinese humanoid robot maker), founded just six months ago, has completed three funding rounds; its MEgo line of embodiment-free capture devices has shipped over 20,000 units in mass production, generating a cumulative 1 million-plus hours of embodiment-free data. Chiron Robotics (Chinese embodied-AI startup) has deployed over 400,000 collection points across more than 100 Chinese cities. MiFeng's chairman and CEO Yao Maoqing put it plainly: embodiment-free data suits pretraining for general representations, but for specific tasks "there's no avoiding real-robot data matched to that embodiment" — internet video distillation is only "a crutch for this stage." Industry standards are also shifting: earlier this year, data captured with phones or action cameras could still be sold, but by the second half of the year, 60FPS, 1080p, and stereo vision have become consensus requirements — "collecting a million hours blindly isn't actually hard; you can get there by repeating a handful of task types, but that kind of data has almost no value." The Ministry of Industry and Information Technology's "Quality Requirements and Evaluation Methods for Embodied Intelligence Datasets — Key Foundational AI Technology," approved in August, takes effect on November 1 and covers eight dimensions including completeness, consistency, diversity, and authenticity. Source: Sina Finance source

Unitree Stock Halves From First-Day Price on 11th Trading Day Since Listing, Market Cap RMB 220.85 Billion · humanoid

Closed at RMB 546.02 on September 2, down 4.39%, exactly half the intraday high of RMB 1,100 on its first day of trading; market cap now RMB 220.85 billion. Shares jumped 460% on their August 19 debut on the STAR Market, at an issue price-to-earnings ratio of 219.23x. On fundamentals, humanoid robots accounted for 51.78% of 2025 revenue, with pure-humanoid shipments exceeding 5,500 units and combined quadruped sales exceeding 33,000 units. Founder Wang Xingxing told investors he hopes investors buy in because they recognize the company's value rather than for speculation, and that going public brings "more pressure but also more motivation." The same day, the company responded to online complaints from employees alleging "penalties but almost no bonuses" and "reimbursements over RMB 100 require Wang Xingxing's personal approval," calling much of the content inaccurate. The company's registered capital has been formally updated post-listing to RMB 404 million. Source: Guangzhou Daily Xinhuacheng source

World's First International Standard for Legged Robots Released, Led by Chinese Experts · humanoid

The standard, titled "Robots — Performance Specifications and Test Methods for Service Robots, Part 5: Legged Robot Locomotion," covers robots that walk, run, climb, and navigate obstacles on legs — including humanoids and quadrupeds — for which there was previously no unified global evaluation framework. The core changes lie in two areas, test environment and evaluation scope: testing is no longer confined to flat-ground mobility but now includes real-world scenarios such as stairs, slopes, and obstacles; evaluation also no longer looks only at motion metrics but factors in perception, interaction, and decision-making as part of task completion. Such robots are already performing real work — performance, inspection, transport, search and rescue — but cross-vendor performance comparisons have so far relied largely on each company's own self-reported demonstrations. Source: Standardization Administration of China source

AgiBot Releases TwinDEX: One Hand Worn by a Human to Collect Data, One Hand Mounted on a Robot to Do the Work · embodied ⚠️ manufacturer's claim

A twinned pair of three-fingered, nine-degree-of-freedom dexterous hands, seven of which are active degrees of freedom. The capture-side unit is a wearable exoskeleton, while the execution-side unit is mounted on a robot; the two are aligned across three dimensions — kinematics (degrees of freedom, joint axes, link proportions), contact mechanics (materials, geometry, surface properties, and tactile sensors), and visual appearance — so that embodiment-free data can be used directly for policy training without needing alignment or fine-tuning against real-robot data. AgiBot's own benchmark claims embodiment-free data collection is 5.3 times more efficient than real-robot teleoperation, and that on multi-task benchmarks, policies trained on either data type improve at the same rate as data volume grows and eventually converge — meaning embodiment-free data can nearly 100% replace real-robot teleoperation data. Both figures come from the manufacturer's own testing and have not yet been independently reproduced. Source: IT Home source

Zeroth Launches Bridge Humanoid Robot and Open-Source OpenBridge Platform · humanoid

Suzhou-based company Zeroth (Chinese robotics startup) has unveiled the compact humanoid Bridge, aimed at developer-focused physical AI development; its hardware, motion-control system, and development interfaces are designed for AI-driven physical interaction rather than preset motion choreography. Out-of-the-box capabilities include stable walking, obstacle avoidance, and a library of humanlike movements, with support for motion capture and VR input. The company simultaneously open-sourced OpenBridge, which includes a public Skill Hub allowing developers to call and extend robot skills written by others. Its Builder program will select 500 developers in its first cohort, with the first 200 receiving a free unit and the remaining 300 gaining priority purchase access to the Geek edition plus SDK access. Founder and CEO Guo Renjie said, "The world never lacks walls; we chose to build bridges instead." Bridge is open for global preorder starting September 2, with the Geek edition's first batch already on sale, priced by region. Source: The AI Insider source

Caterpillar Partners with FieldAI to Bring Robot Foundation Models to Job Sites and Factories · industrial

Caterpillar contributes industry experience, engineering capability, and operational data, while FieldAI contributes embodiment-agnostic autonomy and robot foundation models, built on NVIDIA's technology stack. The initial focus areas are autonomous inspection, digital twins for job sites and factories, early risk identification, and simulation- and automation-driven operational optimization. Chief Technology Officer Jaime Mineart said, "Our industry's future depends on how effectively human expertise and AI-driven machines work together"; John Tuntland, Senior Vice President of the Integrated Components Division, applied the point to Caterpillar's own factories, saying the technologies give teams clearer visibility into how facilities operate and where flow can be optimized. Caterpillar cited FieldAI's track record of deploying in complex, dynamic industrial environments where traditional automation falls short as the reason for the partnership. Source: PR Newswire source