Today's Highlights

· Sereact's dual-arm robots go live in CEVA warehouses in Germany and Poland, sorting Zalando returns

· Waymo pulls safety drivers from vehicles in Detroit; rides will initially be open to employees only

· Singapore's transport minister tells Parliament there will be no cap on the number of autonomous vehicles

· In third-party testing at Mcity, the open-source L4 system Autoware had a collision rate about 1000 times that of human drivers

· Rohit Prasad, former Amazon head scientist for Alexa and AGI, becomes CEO of Boston Dynamics

· Waymo's first debt financing is raised from over $3 billion to $5 billion

Third-party evaluation: open-source L4 system Autoware collides at about 1000 times the human rate · autonomy

The open-source L4 autonomous driving system Autoware.Universe has only 6 of 14 behavioral competencies. Henry X. Liu's team applied the third-party behavioral safety assessment framework proposed in the previous paper to simulation and to real-vehicle testing at the University of Michigan's Mcity test facility. They measured a collision rate of 3.01×10⁻³ per mile, about 1000 times the average for human drivers. The tests also uncovered several previously unknown hazardous scenarios. The previous methodology paper has been cited 5 times.

Henry X. Liu et al. (University of Michigan) · arXiv 2610.08458 source

Levine's group builds a benchmark that reproduces counterintuitive behavior-cloning phenomena in a controlled setting · benchmark

Behavior cloning (BC, learning a policy by directly imitating human demonstrations) often behaves strangely on real robots. The more deeply it overfits the data, the better it performs; fully closed-loop policies without action chunking often fail completely. Seohong Park and Sergey Levine introduce OCBench, which uses scripted policies that mimic the characteristics of human demonstrations to reproduce these phenomena in a controllable way in GPU-accelerated simulation, and uses it to test and overturn several popular explanations.

Seohong Park, Sergey Levine (UC Berkeley) · arXiv 2610.07056 source

JEPA world models "flatten" unstable modes; adding an inverse-dynamics loss preserves them · world-model

When a JEPA world model is trained in the usual way, its loss can drop very low. Meanwhile, unstable modes in the system that can only be suppressed through feedback may be flattened out, so the learned representation cannot be used for stabilizing control. This paper, co-authored by Yann LeCun, adds an action-reconstruction (inverse-dynamics) loss during training and proves that exact reconstruction lets the encoder retain all state directions reachable by actions within H steps. The theory is proven for linear systems, and visual control experiments on CartPole, Walker2D and PointMaze show the conclusion generalizes to nonlinear settings.

Leonardo F. Toso et al. · arXiv 2610.07540 source

OpenWAM: an open, comparable framework for world-action models · world-model

Existing world-action models (which predict future frames and robot actions together) differ in video backbone, interaction structure and supervision, making design choices hard to compare side by side. OpenWAM builds on Wan2.2-5B, first performs causal robot-video pretraining on more than 10,000 hours of video, then attaches an action expert through a shared Mixture-of-Transformers, supporting four generation orders. In the "video first, then action" mode, the success rate on LIBERO-Long rises from 68.4% to 97.8%.

Heng Yu et al. · arXiv 2610.07922 source

A 48.9-million-parameter small policy matches systems more than 40 times larger · manipulation

CoRP uses no VLM and no video-generation prior, and reaches 97.0% on LIBERO, matching systems 40.9 to 163.6 times larger. The authors fix the action generator and swap the visual extractor one component at a time, concluding that performance comes mainly from the visual representation. Pretrained initialization makes or breaks the result; compressing each view to 48 tokens beats keeping all patch tokens by nearly 14 points (97.0% vs. 83.2%).

Nanhe Chen et al. · arXiv 2610.08183 source

QF3: off-policy flow RL trains humanoid locomotion from scratch and transfers zero-shot to hardware · locomotion

The authors say QF3 is the first off-policy flow RL method that can train humanoid locomotion policies from scratch and transfer them zero-shot to hardware. It trains a flow policy with flow matching plus the critic's action gradient, applying gradients only along dimensions close to the replayed actions. When training humanoid locomotion and motion-tracking policies, wall-clock speed is 10 times faster than the on-policy method FPO++.

Chung Min Kim et al. · arXiv 2610.08789 source

Learned error correction for a frozen VLA fails to beat "timed rollback" · vla

On 4 RoboTwin tasks, the authors added a learned error-correction scheme to a frozen π0.5. Across 3888 paired episodes, only 1 task showed a measurable gain (+13.5 percentage points). On that task, simply rerunning failed baseline episodes succeeds 43.2% of the time, versus 63.5% with correction, so many "rescues" actually come from the policy's own randomness. A simple approach that triggers at fixed times and rolls back to earlier joint poses by script showed no measurable difference from learned correction.

Chenchao Sheng et al. · arXiv 2610.06921 source

EgoLAP: turning human-hand video into VLA pretraining data via language-based action reasoning · vla

Human first-person video is cheap, but human and robot bodies differ, so raw trajectories are hard to use directly as control supervision. EgoLAP writes motion intent as structured language actions, paired with reasoning based on scene geometry and object affordances, so that human and robot trajectories share one action chain of thought. Average task progress on real robots reaches 80.1%, 2.3 times that of other action representations.

Lihan Zha et al. · arXiv 2610.08726 source

A physics exam for video world models: best score 57.76 · benchmark

The authors generated 1280 videos with 8 video-generation models, and the top-scoring model reached a total of only 57.76 (out of 100). The benchmark, named World Models' Last Exam in Physics, has 40 controlled tasks covering mechanics, optics, fluids, heat and phase change, electromagnetism and surface tension. Scoring directly measures preset physical relationships for each task and needs no reference video; the evaluator agrees with human judgment more than having a VLM score directly. HF 1↑.

Mingju Gao et al. · arXiv 2610.08791 source

Waymo ｜ First private loan ｜ $5 billion · autonomy

Lenders include PIMCO, Blackstone and Sixth Street, with Goldman Sachs arranging. The size was raised from the over $3 billion originally sought to $5 billion, with the money earmarked for global expansion and rising AI costs. This is Waymo's first time borrowing; it had previously relied on equity financing, and earlier this year raised $16 billion at a $126 billion valuation. Startup Fortune, citing Bloomberg, says the unrated loan is priced more than 500 basis points above the benchmark rate. According to Startup Fortune, Waymo now operates more than 4000 vehicles in 14 US cities, with over 500,000 paid rides per week, and aims to reach 1 million rides per week by year-end.Source: Bloomberg source; Startup Fortune source

Verne (Croatia) ｜ Uber strategic investment ｜ Amount undisclosed · autonomy

On October 1, Verne and Pony.ai (Chinese robotaxi developer) began fully driverless passenger testing in Zagreb; afterwards, Verne co-founder Mate Rimac called Uber an "investor and strategic partner" of Verne on Facebook. When the partnership was announced in March, Uber had only said it "intended" to invest. The post followed a visit to Zagreb by Uber CEO Dara Khosrowshahi, who rode in a robotaxi with no safety driver on board. The division of roles among the three: Pony.ai supplies the autonomous driving system and vehicles, Verne owns and operates the fleet, and Uber provides the ride-hailing network.Source: electrive source; EVWire

Intelligent Machines (US) ｜ Seed round (in talks) ｜ About $100 million · industrial ⚠️ Rumor-level reporting

Former Tesla Optimus AI team lead Ashish Kumar co-founded the company in San Francisco. It builds non-humanoid robots to take on a range of processes in manufacturing and aerospace that normally require skilled workers, mainly using reinforcement learning, and several former Optimus members have joined. According to The Information, the company is seeking a seed round of about $100 million. When Kumar left Tesla in September 2025, he went to Meta.Source: The Information source; Breakingthenews source

Sereact dual-arm robots go live, sorting Zalando returns for CEVA · embodied

Sereact's dual-arm robots have entered live operation in CEVA Logistics warehouses in Germany and Poland, handling returns for Zalando. Returned garments often arrive messily packed, each in a different condition, and cannot be programmed for in advance. Sereact's Cortex platform can recognize, grasp and sort items it has never seen without per-item training, and is charged under a robots-as-a-service (RaaS) model. Zalando co-invested in Sereact's Series B in July, and this go-live is the first deployment milestone of the partnership; Sereact raised a total of $116 million in its Series B this year. CEO Ralf Gulde said: "If a system can only handle products it has seen before, it's simply useless in returns operations. Running in production at CEVA at Zalando's volumes is the strongest proof we have." The press release does not disclose the number of units deployed or daily throughput.Source: GlobeNewswire press release, via The Manila Times source

Waymo pulls safety drivers in Detroit, carrying only employees at first · autonomy

Starting October 6 Eastern Time, Waymo began fully autonomous testing in Detroit with no one in the driver's seat. Rides are open only to employees for now, with public service to follow once the technology is validated. The operating area covers downtown and neighborhoods including Corktown, Midtown, Hamtramck and Indian Village, and will gradually expand to the whole metro area. Waymo has been testing in Detroit with the Zeekr (Chinese EV brand) Ojai vehicle since last November; like Chicago and Montreal, it is one of the northern cities Waymo is trialing to enter regions with variable climates.Source: The Detroit News source

AgiBot's "Aibao" logs about RMB 2.23 million in cumulative sales three days after entering stores · humanoid ⚠️ Single-source reporting

On September 28, 100 units of the Lingxi X2 from AgiBot (Chinese humanoid robot maker) were deployed in 100 stores of ASD (Chinese cookware and appliance maker). According to Shanghai Observer, three days of trial operation produced cumulative sales of about RMB 2.23 million, with sales at the hundred stores rising an average of 39%; the report does not say what period this is compared against. Checkout is still handed off to human staff, and AgiBot itself says "Aibao" is currently "perhaps equivalent to a newly hired sales assistant."Source: World Journal, citing Shanghai Observer source

Kodiak trials fresh produce hauling on California's Highway 99, with a person still in the driver's seat · autonomy

Since September 22, Kodiak AI and Fresno-based carrier DTL Transport have been using autonomous trucks to haul perishable food from Fresno to a distribution center in Los Angeles, with a person still in the driver's seat. Kodiak says these are the first autonomous truck deliveries under California's new rules; it received a heavy-duty autonomous testing permit from the California DMV in August.Source: KMPH Fox 26 source

Autonomous A2Z begins trial operation of an L4 bus in Gwangmyeong, South Korea · autonomy

The GM1 route runs 24 km round trip, connecting Cheolsan Station, KTX Gwangmyeong Station and Gwangmyeong Cave, and is free to ride for the rest of the year; the company announced on the 7th that it has started running. The vehicle is a 14-seat electric bus converted from a Hyundai County EV, fitted with 4 lidars, 8 cameras and 1 radar, with a safety driver taking over in emergencies. The route runs Wednesday through Sunday, with 2 round trips per day. Two intersections also have roadside 128-line lidars that relay information on pedestrians and other objects to the vehicle. The trial runs through December; the company says its cumulative driving distance has reached 1.15 million km.Source: Digital Today source

Boston Dynamics changes leadership: former Amazon AGI head scientist Rohit Prasad becomes CEO · humanoid

Rohit Prasad took over Boston Dynamics on October 7 and will also join the board once approved through the due process. He spent 12 years at Amazon, last serving as senior vice president and head scientist for Alexa and AGI. He took Alexa from early R&D to use in hundreds of millions of homes, and in recent years led Amazon's Nova foundation model family; before that, he spent nearly 14 years on machine learning research at Raytheon BBN. The company's announcement says the appointment is meant to "combine world-class robotics with advanced AI capabilities to commercialize intelligent machines at scale." Jaehoon Chang, vice chairman of Hyundai Motor Group and chair of Boston Dynamics' board, said: "Boston Dynamics' industry-leading robotics, Rohit's proven ability to turn AI into products at global scale, and Hyundai Motor Group's manufacturing, logistics and mobility capabilities give us a solid foundation to build and scale Physical AI."Source: The AI Insider source

Singapore: no cap on the number of autonomous vehicles · autonomy

Singapore's Minister for Transport Jeffrey Siow said in a reply in Parliament on October 7 that, given that transport demand is still growing, the government will not cap the number of autonomous vehicles. He estimates that by the end of 2028 Singapore may have several hundred autonomous passenger vehicles, against about 72,000 currently active taxis and private-hire cars. Punggol currently has 15 driverless vehicles, and Grab plans to expand to 50 within the next six months. Siow said: "While it is impossible to predict how many autonomous vehicles Singapore will have in five years, it is almost certain that the majority of point-to-point transport services will still be provided by human drivers." On supporting measures, two job-transition programs provide 350 positions, and the government gives participating employers wage subsidies of up to 90%; operators must report accidents involving third parties or property damage. Waymo plans to test in Singapore in 2027 and open bookings in 2028.Source: The Business Times source

South Korea launches a national logistics-humanoid project; Lotte Hi-Mart plans to put humanoids in warehouses in 2028 · humanoid

KEIT, under South Korea's Ministry of Trade, Industry and Energy, has selected a logistics-humanoid project under the "2026 Robot Industry Technology Development Program," running to March 2031, with the goal of using Korean-developed Physical AI models to reduce reliance on foreign robots and AI models. According to Asia Economy, the project is led by Yujin Robot, with the Korea Electronics Technology Institute participating and Lotte Global Logistics and Lotte Hi-Mart joining as demand-side partners. Lotte Global Logistics will pilot 6 processes at its Jincheon, Yangsan and Icheon logistics centers, including 3D sorter operations, GTP station pick-and-place, and consolidated packing of small items, requiring humanoids to complete the entire flow from confirming an order to loading items into the sorter; last year it already trialed a flow from garment picking to bagging with a bipedal humanoid at Jincheon. Lotte Hi-Mart plans to deploy humanoids at its Icheon logistics center from 2028, and will first set evaluation standards in a simulated site by the end of next year.Source: Chosunbiz source; Asia Economy source

Lisa Su meets Samsung's Jun Young-hyun; sources say talks extended to Physical AI · hardware ⚠️ Rumor-level reporting

AMD CEO Lisa Su had dinner in Seoul on the evening of October 7 with Jun Young-hyun, vice chairman of Samsung Electronics' DS division, talking for about 3 hours, their first meeting in 7 months. Citing people familiar with the matter, Seoul Economic Daily reports that the two sides discussed expanding HBM4 supply for AMD's next-generation AI chip Instinct MI455X, commissioning Samsung for foundry work, and extending the partnership to Physical AI. AMD acquired World Labs for $8.2 billion at the end of last month; Samsung is preparing a next-generation robot chip solution that integrates memory, image sensors, foundry and advanced packaging. Asked about Physical AI cooperation, Su only said the two companies "will cooperate in multiple areas going forward."Source: Seoul Economic Daily source

Beijing forms a robot insurance co-insurance pool, and dexterous hands get a policy too · adjacent

The dexterous hand Y-Hand M2 signed an insurance policy at the China International Fair for Trade in Services (CIFTIS), which closed on September 13. Its developer Yue Quan Bionics signed with China Pacific Property Insurance (CPIC unit), with coverage including damage from accidents and natural disasters, losses from defects in design, installation materials or operational errors, and third-party bodily injury and property damage during use. CPIC's Beijing branch says this is an industry first for a key component of embodied robots. Last month, the Beijing branch of PICC Property & Casualty (Chinese state insurer) and several other insurers formed the Beijing Intelligent Robot Insurance Co-insurance Pool, whose first product covers equipment failures after the original manufacturer warranty expires. Chen Jian, head of the technology insurance sub-center at PICC P&C's Beijing branch, said that core precision components in robots wear out quickly and manufacturer warranty periods are limited; "once equipment is out of warranty, repairs and part replacements are very expensive, which to some extent constrains industry expansion and commercialization."Source: China Daily source

Tesla to show the Cybercab at the Paris Motor Show · autonomy

This will be the Cybercab's first public showing in Europe. The EU vote on FSD (Supervised) has been postponed and will be taken up again in December at the earliest.Source: Bloomberg source; Automotive News source

Incheon builds a Physical AI test center, opening to companies in April 2027 · industrial

The city of Incheon is setting up a Physical AI test center at Incheon Global Campus, split into two zones: a real-scene simulation zone that uses AMRs and collaborative robots to recreate factory and warehouse environments, and a virtual testing zone running on GPU servers. Design and infrastructure work will finish within October, with support for companies starting in April 2027; the city then plans to connect Incheon Airport, Incheon Port and national industrial complexes for larger-scale testing.Source: Seoul Economic Daily source