Today's Highlights

· Boston Dynamics opened an Atlas training center at Hyundai's Georgia plant, with humanoid robots beginning to practice parts sequencing on a real automotive production line

· Cognex acquired RealSense for $500 million in cash, with total consideration around $600 million including retention incentives

· Tesla's fleet actually carrying paying riders in Austin over the past 7 days fell back to 8 vehicles, even as state-registered Cybercabs rose to 67

· Google's Intrinsic open-sourced the control, motion planning, and simulation modules of Intrinsic Core under an Apache 2.0 license

· GPT-6 Astra, used directly as a robot policy, achieved a 22.48% average success rate across 42 RoboDojo tasks, ranking above 40 published policies

RoboDawn: Closed-loop robotic arm control directly from a VLM, no robot training data required · vla

Today's most-discussed paper in the community (HF 96↑). The authors decompose robot control into a set of discrete translation, rotation, and gripper commands, which an agentic VLM observes, reasons about, executes, and adjusts based on results — all without any task-specific robot training. On RoboTwin 2.0 C2R, zero-shot success rate is 53.2%; adding one in-context demonstration raises it to 73.6%, above π0.5's 46.0%. The method was also validated on a real Franka arm across two tasks: placing blocks into a basket and stacking blocks.

Meng-Hao Guo et al. · arXiv 2609.22966 source

First systematic evaluation of GPT-6 Astra on RoboDojo · vla

Following last week's RoboHarm and dual-arm YAM tests, this paper puts "LLM as a direct policy" against a full benchmark. GPT-6 Astra ran 2,100 trials across all 42 RoboDojo tasks, achieving a 22.48% average success rate, ranking above all 40 published policies. Using the same post-processing, GPT-5.5 managed only 0.88%. Its capability profile is highly polarized: it generalizes well on tasks requiring semantic understanding, but performs poorly on tasks requiring high precision, dynamic control, or complex bimanual coordination — and providing one demonstration did not improve overall performance.

Wenbo Zhang et al. · arXiv 2609.24170 source

Grounded Action Model: a robot foundation model built on 3D grounding · vla

Existing pretrained backbones for VLA and world-action models don't enforce explicit learning of "where the target object is and how large it is" — this paper puts metric geometry at the very front of the model (HF 18↑). Language, point, or box prompts are first converted into object-centric representations, then combined with state history to predict action chunks. Under 16 types of perturbation on LIBERO-PRO, average success rate is 61%, versus 53% for π0.5. On a real dual-arm YAM setup under visual shift, it succeeded 17/20 times versus π0.5's 4/20.

Gehao Zhang et al. · arXiv 2609.23863 source

CARE: a VLA execution framework that learns error correction from real failures · manipulation

Error-correction data is usually generated via artificially designed perturbations; CARE instead collects trajectories from real policy failures, statistically models the deviation distribution following failure by task stage, and uses this to synthesize failure states and corrective demonstrations (HF 10↑). It proves effective across multiple VLA backbones: average simulation success rate improves by 14.5 points, and dual-arm real-robot success improves by 15.9 points. Code, models, and data are public, along with a companion failure-recovery benchmark, FSR-Bench.

Junlan Xiao et al. · arXiv 2609.24118 source

EgoWild2Dex: teaching dexterous hands from first-person video shot in homes, factories, and pharmacies · manipulation

Human video datasets are typically collected in purpose-built settings; this paper instead has people wear head-mounted cameras during everyday work, compiling the EgoWild dataset totaling 538.9 hours: 179,049 clips covering 1,282 object categories. The accompanying method aligns the shaky human egocentric viewpoint with robot observations. Across three long-horizon dual-arm dexterous tasks, average real-robot success rate is 96.7%, with 33.3% zero-shot success on novel objects. The authors state that data, models, and code will be released.

Kunyang Lin et al. · arXiv 2609.23755 source

RAPolicy: 1-2 hour on-robot online post-training for VLA · vla

Sampling and learning run asynchronously in parallel, with both critic and actor updates anchored to replay data, reducing reliance on value estimates for out-of-distribution actions. Each task starts with only 10 demonstrations and 1-2 hours of online training. Average success rate across four single-task settings is 86.3%; in a five-task joint setting it rises from 52% to 88%, with less human intervention required.

Jiarui Yang et al. · arXiv 2609.22888 source

Whether actions are written as absolute values or deltas gives a world model two different answers · world-model

For the same instruction trajectory, writing it as absolute joint targets versus relative deltas breaks a latent dynamics model trained on the other encoding: retrieval performance drops by 2.6 to 13.4x across three robot datasets, and goal-conditioned action selection falls from 53% to 15%. The two encodings can be reconstructed from each other (R² 0.996), so no information is actually lost — the problem is that the action channel was never audited. The authors' proposed fix is to average the training objective across both encodings and add a divergence penalty to cover the worst case.

Ahmed Karim et al. · arXiv 2609.23252 source

Uranus: replacing hand-built robot simulators with an autoregressive diffusion model · world-model

Given a future joint-position trajectory as input, the model generates multi-view footage step by step, with no fixed duration limit, reaching 24 FPS after inference optimization, and supporting multiple robot embodiments and camera configurations. Code and weights are released, and the paper separately lists current limitations.

Wenkang Qin et al. · arXiv 2609.24815 source

Cognex × RealSense | Acquisition | $500 million cash | Total consideration ~$600 million · hardware

Machine vision maker Cognex is acquiring depth-camera company RealSense using its own cash, at a base price of $500 million, with closing expected in the fourth quarter. Israeli outlet Globes' reported figure of "over $600 million" includes two additional components: a three-year cash retention plan worth $56.5 million, and roughly $50 million in restricted stock. RealSense's quarterly revenue was only $7-8 million when it spun off from Intel two years ago; Cognex projects it will generate $80-90 million in revenue in 2026, up more than 50% year-over-year. Intel retains a 20% stake and will exit alongside the deal.Source: Globes source, Cognex investor announcement

New venture from Chen Long (former head of Xiaomi's XLA program) | Funding secured | Amount undisclosed · world-model ⚠️ Single-source

Chen Long previously led the foundation model for Xiaomi's automotive driver-assistance system and its XLA large model program; he has now moved to Silicon Valley to start a company as co-founder, partnering with a former Google DeepMind scientist to build an embodied foundation model. 21 Jingji (Chinese financial outlet) reports the venture has already secured funding and commercial partnerships. One investor said Chen Long is aiming to build a "self-evolving" embodied brain, rather than chasing scaling laws through more data or modified training methods. Lei Jun (Xiaomi founder) has expressed support for the venture. Xiaomi's driver-assistance team has since brought in several executives from Horizon Robotics (Chinese AI chip company) to take over the work of Chen Long and former L3 lead Wang Naiyan.Source: 21 Jingji source

Xiantu Origin | Seed round | Tens of millions of yuan · embodied

Led by Qifu Capital (Chinese venture firm). The company was registered in June this year; its core team comes from the National University of Defense Technology's robotics lab (which produced China's first humanoid robot, "Xianxingzhe"). It is positioned to perform fine-manipulation embodied tasks on industrial sites.Source: DoNews source

Camsense | Hong Kong IPO offering | ~$78 million · hardware ⚠️ Single-source

Following its listing hearing approval on September 13, this maker of sensor components for robot vacuums has begun its share offering.Source: The Bamboo Works source

Atlas moves into Hyundai's Georgia plant, begins practicing parts sequencing on a real production line · humanoid

Boston Dynamics has opened a Robotics Metaplant Application Center (RMAC) inside Hyundai Motor Group's U.S. plant, HMGMA. Located near Savannah, the plant produces electric and hybrid models including the Ioniq 5 and Ioniq 9. There, Atlas is working on pre-assembly parts logistics and sequencing — arranging parts in the order needed for assembly. Next year the training center will move into a new building roughly 10 times larger, and will begin exploring industries beyond automotive. Hyundai's goal is to deploy 25,000 Atlas units across Hyundai and Kia's global plants over the next several years. For now this remains a training and testing phase, still some distance from stable-cadence mass production work.Source: The Korea Times source

Tesla's actual paid-rider fleet in Austin falls back to 8 vehicles · autonomy ⚠️ Compiled data

Electrek, citing data from community tracking site Robotaxi Tracker, reports that only 8 Tesla Robotaxis were observed carrying paying riders in Austin over the past 7 days — similar to levels seen when the service first launched in June 2025. Around the September 3 Cybercab launch event, the figure briefly surged past 100 vehicles, but has sat between 20 and 40 for most of the year. Over the same period, Tesla registered 28 more autonomous vehicles with Texas's DMV, bringing the state total to 476, including 67 Cybercabs and 409 Model Ys. Registration counts don't reflect actual ridership, and the gap between the two figures is widening.Source: Electrek source, eletric-vehicles.com source

LS Electric moves two Spot units to routine inspection duty · industrial

The two Boston Dynamics Spot robots are deployed at LS Electric's Busan plant, its main production site for ultra-high-voltage transformers. Following roughly two months of pilot operation with Hyundai AutoEver starting in July, they have now gone into official use. Spot uses thermal imaging for non-contact inspection of gas-drying chamber door seals to detect heat leaks, while also collecting thermal, noise, and vibration data from pumps, motors, and control cabinets to give early warning of faults. LS Electric plans to use this data to train an AI analysis model, then extend it to other plants at home and abroad.Source: The Korea Times source

Over 2,000 humanoid robots already deployed in Guangdong factories · humanoid

At the "AI + Manufacturing" session of Guangdong's second AI Application Conference, 8 large-scale embodied robotics application projects were signed, totaling 75 million yuan in investment, spanning production lines, libraries, subways, communities, and sports venues. The conference disclosed that Guangdong manufacturers have already deployed over 2,000 humanoid robots, with the total expected to surpass 3,000 by year-end.Source: 21 Jingji source

Aurora sets 2028 target for positive free cash flow · autonomy ⚠️ Company projection

Aurora currently operates 20 driverless heavy trucks on public roads, running the Houston-Dallas and Fort Worth-El Paso routes, with a goal of reaching 200 vehicles by year-end. The founder and CEO told CNBC that fleet expansion would make the company free-cash-flow positive by 2028. For comparison, Aurora lost more than $800 million in 2025.Source: CNBC source

XPeng begins rolling out new version of its second-generation VLA · autonomy ⚠️ Manufacturer claim

XOS 6.3.0 began rolling out on September 22. The distilled VLA Lite version debuts on the G9L Max, and the update also covers the MONA L03, GX, G7, P7, and the 2026 M03, G6, G9, P7+, and X9.Source: Autohome source

XPeng Robotics completes supplier audits and core component sourcing decisions, with suppliers heavily overlapping Tesla's · humanoid

On September 22, XPeng Group held its first Robotics Supply Chain Partner Conference, attended by suppliers of actuators, dexterous hands, sensors, AI chips, and other components. According to The Paper (Chinese news outlet), XPeng Robotics has completed supply chain audits and finalized sourcing decisions for core components, with its technical roadmap, product form, and mass-production timeline all now defined. Just last week, Tesla engineers were in China auditing Optimus suppliers; sources say the two companies show a "high degree of overlap" at the core supplier level.Source: The Paper source

Mercedes-Benz signs with Wayve to put AI Driver into production vehicles within about two years · autonomy

Mercedes-Benz has signed a multi-year agreement with UK-based Wayve to integrate Wayve's end-to-end autonomous driving system into future models, based on Mercedes' MB.OS architecture, within roughly two years. Mercedes previously participated in Wayve's Series D funding round.Source: Global Banking & Finance Review source

Figure Helix 2.5: zero-shot household chores in 30 unfamiliar homes, 56% success rate · embodied ⚠️ Company claim

Figure 03 performed three household chores — including picking up scattered toys into a basket — across 30 homes in the Bay Area it had never visited, with no on-site fine-tuning, achieving a 56% success rate. With task data, architecture, and evaluation held constant, pretraining alone on the human-behavior dataset Index raised zero-shot success from 9% to 56%. Figure claims that each time it doubles pretraining data, loss decreases smoothly enough to be predicted in advance, and asserts this shows it has found a scaling law for robotics. Viewed another way, nearly half of attempts still fail.Source: Robotics Frontier (Chinese industry outlet) source

2025 global humanoid robot shipments: IDC says 18,000 units, IFR says about 7,000 · adjacent

TechJuice, citing IDC data, reports global humanoid robot shipments of 18,000 units in 2025, up 508% year-over-year, with Chinese manufacturers accounting for 90% of the total. Yesterday IFR gave a figure of about 7,000 units, with its count limited to industrial and professional service applications. The two organizations use different counting methodologies, resulting in figures that differ by more than double.Source: TechJuice source

Epson to invest ¥280 billion over three years in robotics and precision businesses · industrial ⚠️ Company plan

President Yoshida Junkichi announced in Seoul that Seiko Epson will invest ¥280 billion (about 2.4 trillion won) over the next three years, equivalent to 20% of last year's ¥1.4 trillion in revenue, with a focus on industrial robots and precision components. The goal is to raise new businesses' share of revenue from 33% last year to 50% by 2035. The same day, Epson launched its first collaborative robot.Source: Seoul Economic Daily source

Industry standard for embodied dataset quality takes effect November 1 · adjacent

The industry standard YD/T 6771-2026, led by China Academy of Information and Communications Technology (CAICT) and drafted with over 40 organizations, evaluates embodied dataset quality across dimensions including completeness, consistency, diversity, authenticity, and safety. As of the end of June, China had over 70 embodied AI training facilities already operational, with more than 40 others under construction or planned.Source: Pandaily source

Yangtze River Delta exports 16 billion yuan of four key robot categories in first 8 months · industrial

Exports of four key robot categories, including surgical robots, grew 3x year-over-year.Source: Guandian.cn source