Today's Highlights

· Pony.ai (Chinese robotaxi developer) and Verne have begun fully driverless passenger testing in Zagreb, a first for Europe, on a 22 km route running straight to the airport

· NHTSA has linked 5 crashes and 3 deaths to comma.ai's aftermarket driver-assistance kit, which is installed in more than 30,000 vehicles

· IronMind raised human first-person-video pretraining to 10,000 hours, lifting the success rate on new real-world humanoid tasks from under 12% to 55%

· World-model planning has a hard ceiling: even with perfect prediction, moving the goal from 5 steps away to 20 steps away drops success from 92% to 41%

· Figure had a retired F.02 jump into molten steel; this batch of robots moved more than 90,000 parts on BMW's production line

· Anthropic estimates robots can already do 74% of physical tasks in the US, but only 0.3% are cheaper than humans on a cost basis

IronMind: Humanoid dexterous manipulation only leaps once human first-person video reaches 10,000 hours · vla

Human first-person video is plentiful and cheap to collect, but human hands and robot hands differ in structure, so training humanoids on it has always suffered from an embodiment gap. IronMind uses a different action representation: actions are defined directly in the camera frame, skipping the body-retargeting step. Pretraining data grew from 250 hours to 10,000 hours, and validation loss fell roughly log-linearly with data volume. On six real-robot tasks with unseen objects, the 10,000-hour model reached a success rate of 55.0%; pretraining tiers of 5,000 hours and below topped out at 11.7%, and no pretraining scored 5.0%. The jump from 5,000 to 10,000 hours is much steeper than the smooth decline across the earlier tiers.

Huimin Pan et al. · arXiv 2609.39403 source

Planning limits of latent-space world models: they only see clearly 5 to 10 steps ahead · world-model

This paper specifically tests where world models fail. The authors trained action-conditioned predictors on five frozen vision backbones, including V-JEPA 2 and DINOv2. The models could reliably rank actions only for goals within 5 to 10 steps, while task goals are typically 16 to 53 steps away. Scaling the predictor up 81 times did not extend this range. Switching to a real simulator for perfect prediction, moving the goal from 5 steps away to 20 steps away still dropped success from 92% to 41%. Within the plannable range, using it to rerank 8 candidate actions from a VLA raised success on 16 tasks from 65% to 77%.

Ali Alrasheed et al. · arXiv 2609.39235 source

Galbot (Chinese embodied-AI robotics firm) evaluates GPT-6 Astra as an embodied policy: useful for decisions, too slow for control · vla

Earlier issues covered tests of Astra directly controlling robot arms and navigating; this paper tests six domains together and also tallies the cost. Astra mixed with π0.5 for control reached 48% success on a RoboDojo subset; 92% on RxR instruction navigation; and in legged obstacle crossing, all 5 attempts failed to reach the goal. Across 50 RoboDojo instances, direct control consumed 1.132 billion tokens in total. A 30-second legged motion required 250 model calls averaging 39.86 seconds each, during which the physics simulation had to be paused. HF 14↑.

Xuchuan Chen et al. (Galbot) · arXiv 2609.38537 source

DrivingBench: letting a general-purpose model drive a real Toyota Corolla · autonomy

This is the first benchmark to have a general vision-language model drive a real car. The model watches the Corolla's camera feed and weaves slowly around cones in a parking lot. The car does not stop while the model thinks, so inference latency counts against it. The models tested were GPT-6 Astra, Claude Fable 5.1, GPT-5.6 Sol and Grok 4.6, each with up to three attempts. Only Astra, on its second attempt, completed the full course; none of the others got past halfway on any attempt. The authors have released the test framework, prompts and complete logs with video telemetry.

Aditya Ramabadran et al. · arXiv 2609.38948 source

Counterfactual training (ECT): fixing VLAs that "hear the instruction but replay the old trajectory" · vla

VLAs can exceed 90% success within the training distribution, yet fail when the scene changes and a different action is needed. The authors call this a failure of "instruction-action binding": π0.5 and GR00T-N1.7, on hearing a familiar instruction, directly recall the trajectory seen during training. The remedy is to construct demonstrations in pairs, where the same instruction maps to different actions in distinguishable scenes, and train on both demonstrations in the same update. On a real UR5e, with the number of demonstrations unchanged, success at unseen positions rose from 8% to 88%.

Hung-Jen Chen et al. · arXiv 2609.39971 source

BIND: using camera geometry to bind 3D actions to 2D image features · manipulation

Most existing policies' action heads regress coordinates from a single global feature vector, which tends to break when the camera position shifts slightly. BIND projects candidate end-effector positions back into each view, binds them to pretrained features there, and then scores them; this correspondence comes directly from geometry and does not need to be learned from demonstrations. On a real robot, 5 demonstrations suffice for near-complete success; when the camera viewpoint shifts substantially, the coordinate-regression baseline fails entirely while BIND's performance degrades only gradually.

Cameron Smith et al. · arXiv 2609.38443 source

SimEX: a coding agent practices in simulation first, then calibrates with 10 minutes on a real robot · manipulation

A coding agent first iterates by trial and error in simulation to build up a robot toolbox. Once on the real robot, each trial is used to correct the simulator, and the corrected simulator is then used to diagnose failures and screen candidate fixes. On real tasks such as folding towels, scanning barcodes and handling plates, it learned the skills with no demonstrations and only 10 minutes of real-robot interaction.

Jiaheng Hu et al. (including Peter Stone, Roberto Martín-Martín) · arXiv 2609.38982 source

WorldAuditBench: having multimodal agents find clipping and floating objects in 3D worlds · benchmark

Simulated worlds are increasingly used to train robots, and spotting glitches such as floating objects or walls that can be walked through requires an agent to look around as it moves. The benchmark contains 13 environments and 213 anomaly tasks. Five frontier models scored between 6.6% and 42.3%, versus 83.4% for humans. HF 66↑, the most popular paper in the community today.

Ziyan Jiang et al. · arXiv 2609.40325 source

FORT Robotics ｜ SPAC merger listing ｜ expected to raise about $201 million ｜ enterprise value $556.6 million · adjacent

FORT and Newbury Street II have confidentially submitted a draft S-4 registration statement to the US SEC; after the merger the company plans to list on Nasdaq under the ticker FROB. FORT works on the robot safety communications and emergency-stop layer, with customers including Agility, DoorDash, Zoox and Google DeepMind. Revenue in 2025 grew 62% year on year, with no single customer above 9%. The roughly $201 million gross proceeds assume zero redemptions by public shareholders, and the deal is expected to close between Q4 2026 and Q1 2027. The same week, Agility signed a memorandum of understanding with FORT, under which FORT will supply Digit 5 with three parts: an emergency-stop pendant, onboard communications and an external safety interface. At a pre-money equity value of $500 million, the market is pricing FORT on its position in the safety-standards layer, out of proportion to its current revenue scale.Source: scanx.trade source; PR Newswire source

Yiwei Technology (Zhuhai; Chinese robotics SoC developer) ｜ Hong Kong IPO ｜ passes listing hearing · hardware

After its Hong Kong listing filing cleared in mid-September, Yiwei Technology disclosed its post-hearing information pack on September 30 and has passed the Hong Kong Stock Exchange listing hearing; its joint sponsors are CSCI (China Securities International) and Guotai Junan Capital. The company builds intelligent mobile robot solutions using in-house robot algorithms plus a dedicated SoC. According to Frost & Sullivan data, it ranks in the top three in China's intelligent mobile robot solutions market by 2025 revenue.Source: Shenzhen Commercial Daily · Duchuang source

Destro AI ｜ Seed round ｜ $8 million · industrial

Destro came out of stealth this Tuesday. It does not build robots; it builds a warehouse orchestration layer that assigns work to both robots and on-site workers. The first deployment is cross-dock sorting for Yusen Logistics' US operations. Founder Manthan Pawar said: "One of the biggest reasons we win against robotics companies is that we're not a robotics company." He said the company is on track to become cash-flow positive by the end of this year.Source: TechCrunch source

Tujian (途见科技; Chinese e-skin startup) ｜ Pre-A++ round ｜ hundreds of millions of yuan · hardware

Co-led by the Beijing Artificial Intelligence Industry Investment Fund and the Beijing New Materials Industry Investment Fund. Tujian makes stretchable flexible electronic skin, entering the market through tactile data gloves and first-person data-collection solutions; its material stretches by more than 100%. Dealroom records the round as $15 million.Source: 36Kr; Dealroom source

Pony.ai and Verne launch Europe's first fully driverless passenger testing in Zagreb · autonomy

The test route is 22 km, connecting Verne's headquarters, a major business district in Zagreb and Franjo Tuđman Airport, and will expand to the entire operating area over the coming months. The service went live in early April, at which time vehicles carried safety operators. Since then the fleet has driven more than 200,000 km, completed thousands of passenger trips, and earned an average passenger rating of 4.7. The division of labor: Pony.ai supplies its seventh-generation Robotaxi and the autonomous driving system, Verne handles operations, and Uber provides the ride-hailing entry point. Pony.ai CEO Peng Jun said this step shows its "Virtual Driver" can be combined with local operating capability to form a scalable commercial path. It is still test passenger service, not yet fully driverless paid operation.Source: Robotics & Automation News source

Kodiak lands its first fully driverless long-haul customer: IKEA, 219 miles on I-45 · autonomy ⚠️ Company-stated plan

Kodiak announced on September 29 that IKEA will be its first partner for cab-empty operations, starting later this year. The first 219-mile segment mainly runs along I-45 between Houston and Dallas, part of IKEA's 292-mile route from its Baytown distribution center to its Frisco store. Over the past four years, Kodiak trucks with safety drivers aboard have hauled more than 1,300 loads for IKEA, covering 750,000 miles in total. Highway safety validation must still be completed before launch, so the start date remains the company's plan.Source: CDLLife source

Northrop Grumman uses cobots to cut one F-35 avionics assembly step from hours to minutes · industrial

The step is bolting the front housing of the F-35's Communications, Navigation and Identification (CNI) module. CNI line manufacturing manager Lavell Catlin said the cobots "cut work that used to take hours down to just minutes," and freed up several employees for other work. The robots work alongside technicians and do not complete the assembly on their own. The company's integrated assembly line in Palmdale turns out one F-35 center fuselage every 30 hours.Source: Defence Industry Europe source

Momenta (Chinese autonomous-driving company) says its Robotaxi fleet will reach "thousands" by end-2027 · autonomy ⚠️ Company-stated plan

Finimize, citing Reuters: Xie Shuo, head of Momenta's Robotaxi business, said the company currently has more than 100 vehicles deployed in three countries, will reach several hundred by the end of this year and thousands by the end of 2027, and will expand to Dubai next year. Part of the cost reduction comes from developing the next-generation X9 driver-assistance chip with chipmaker XHeart.Source: Finimize source

Comau validates a collaborative picking and palletizing solution for Decathlon's e-commerce warehouse · industrial

The solution uses Comau's MyCo cobot with modular grippers and vision recognition, and was tested and validated in Decathlon's e-commerce fulfillment operations, as part of the EU's MASTERLY project. It is currently at the project validation stage, and no plans for large-scale deployment have been announced.Source: Robotics & Automation News source

Figure has a retired F.02 jump into molten steel; F.03 takes over · humanoid

At BMW's Spartanburg plant, the F.02 loaded more than 90,000 parts and took part in producing over 30,000 X3s. The F.03 now takes over. Figure's reasoning: selling off the old fleet could leak technology, and dismantling units one by one would slow down the F.04. CEO Brett Adcock asked for disposal ideas on X, and Arnold Schwarzenegger replied "melt them." Foundries in the US and Mexico would not let robots with lithium batteries jump into their equipment, and in the end a foundry in Imatra, Finland agreed. The robot used a simulation-trained jumping model, modeled on stuntman movements, and leapt into a 75-ton electric arc furnace. Humanoid fleets are starting to be retired at scale, and how to scrap and recycle them is a question nobody has needed to answer before.Source: Interesting Engineering source

Maxwell, from a Chinese Academy of Sciences team, tops the Meta-World leaderboard · world-model ⚠️ Single-source claim

According to SCMP, Maxwell, an embodied model developed by the Institute of Artificial Intelligence for Industries under the Chinese Academy of Sciences, scored 91.9 on the Meta-World simulation benchmark, ranking first. Second is FabriVLA from Youibot (Chinese mobile robot maker) at 90, and third is SUREFlow from South Korea's Kyungpook National University at 88.3. Meta-World has 50 tabletop manipulation tasks, and the top three are within 4 points of each other, close to saturation. The world-model planning limits study in today's papers section also uses this benchmark.Source: South China Morning Post source

Italy's Generative Bionics signs with Eni; humanoid welder to be first deployed in early 2027 · humanoid

This company, spun out of the Italian Institute of Technology, closed a €70 million round last December led by CDP Venture Capital, and unveiled Gene.01 in July. Its approach is one general-purpose platform customized per task; CEO Daniele Pucci's line is "a robot that does everything does nothing well." The company has a four-year project with shipbuilder Fincantieri to develop a humanoid welder, with first deployment planned for early 2027. The newly signed Eni agreement will first explore inspection and teleoperation at plant sites, along with joint R&D with Eni's Versalis on lighter foot materials.Source: Il Sole 24 Ore source

XPeng (Chinese EV maker) brings its IRON humanoid and the G9L to the Paris Motor Show · autonomy

The Paris Motor Show runs October 12 to 18. XPeng will give the G9L its global debut there, and its stand will also show the IRON humanoid robot and a flying car. Only two automakers, XPeng and Tesla, are taking part in the show's official autonomous driving experience; XPeng will use the L03 for NGP ride-alongs, its first large-scale overseas ride-along program.Source: CleanTechnica source; eletric-vehicles.com source

Estun (Chinese industrial robot maker) says it ranked first in China's industrial robot shipments in H1 · industrial

At an investor relations event, Estun cited MIR DATABANK data: the company ranked first in China's industrial robot shipments across all brands in the first half, has led for six consecutive quarters, and is the first Chinese brand to exceed 10,000 units in quarterly sales. Heavy-payload robots and SCARA robots are growing faster.Source: Southern Finance source

A SoftBank, AIST and Mitsubishi Electric team wins Japan's robot foundation model competition · world-model

The competition was organized by the AI Robot Association of Japan (AIRoA) and ran from April 2025 to February 2026, with the task of developing foundation models for service robots in daily life.Source: SoftBank source

Hyundai-Kia and the city of Seoul sign to advance L4 urban buses · autonomy

What was signed is a memorandum of understanding. Hyundai-Kia will develop autonomous buses with standardized sensors and redundant components, plus a real-time control platform; the city of Seoul will handle policy and data sharing.Source: National Business Daily source

① NHTSA opens an investigation into comma.ai's aftermarket driver-assistance kit · autonomy

On September 29, NHTSA announced an investigation into comma's aftermarket "hands-free" driving kit. It links 5 crashes to the device, two of which reportedly caused 3 deaths in total. The kit is installed in more than 30,000 vehicles, mostly older models. Products like this plug into the original vehicle's systems and fall outside the existing certification chains of both automakers and regulators; what this investigation must answer is who is responsible for that gap.Source: TechRadar; Repairer Driven News

① The US mobile robot safety standard R15.08 is now complete, with all three parts out · industrial

A3 has published ANSI/A3 R15.08-3-2026, completing the final part of the series. Parts 1 and 2 cover the robot itself and system integration; Part 3 covers the user: it requires companies to conduct risk assessments, continuously maintain and document the risk-reduction measures the assessment identifies, and follow change management when the work environment changes. Responsibility extends from manufacturers and integrators to users.Source: Automation World source

① Minneapolis votes on whether driverless ride-hailing needs a human monitor · autonomy

The city council is set to vote on October 1 local time on an ordinance requiring a human monitor in driverless ride-hailing vehicles. The ordinance passed the committee of the whole 7 to 4 on Tuesday and will most likely pass, but Mayor Jacob Frey has consistently criticized the measure and may veto it. Waymo is already testing in the Twin Cities area. This spring, the Minnesota legislature agreed on a bipartisan basis that statewide legislation was needed, but it ultimately failed to pass.Source: MPR News source

② Anthropic: robots can do 74% of physical tasks, but only 0.3% are cost-effective · adjacent

Anthropic published research on September 30, with estimates based on about 900 O*NET occupations and about 19,000 tasks. The conclusion: existing robot technology can technically perform 74% of physical tasks in the US, corresponding to 34% of working hours; only 0.3% are cheaper than humans on a cost basis. Robot prices have fallen about 3% a year since the 1990s, and at that rate it would take about 40 years for that share to reach 10%. As an example, packers: the US has 560,000 of them, with a median annual wage of $49,000, and robots taking over entirely would cost $45,400 a year, already cheaper; a dishwashing robot costs more than three times as much as human labor. The score and cost for each task were estimated by Claude, and the authors backtested using labor market data from the past 50 years.Source: Anthropic; GIGAZINE source

② JPMorgan: 1.75 million humanoid and legged units shipped globally in 2030; average price down a cumulative 56% by 2035 · humanoid ⚠️ Estimate

A JPMorgan report forecasts combined global humanoid and legged robot shipments rising from about 60,000 units in 2026 to about 1.75 million units in 2030, with China going from about 50,000 to 900,000 over the same period. To drive volume, the blended global robot average price falls by a cumulative 56% by 2035, so shipment growth will not convert proportionally into revenue and profit. The report sorts targets into three groups: automaker-affiliated, Tesla supply chain, and robotics-native startups, with fiscal-2027 price-to-sales ratios ranging from under 1x to over 45x, and a tilt toward upstream core components. The 60,000 here includes legged robots, so it cannot be compared directly with the humanoid-only H1 figures from IDC and Sullivan reported last week.Source: Caiwen source

② IFR service robot report: nearly 250,000 professional service robots sold in 2025 · industrial

The IFR's World Robotics 2026 Service Robots shows professional service robot sales grew 24% in 2025, to nearly 250,000 units. Transportation and logistics is the largest segment, with 117,500 units sold, 47% of the total. Consumer service robots reached 34.2 million units, up 37%; surgical robots reached 8,400 units, up 57%. The IFR cautions that the data comes from a sample of 238 suppliers and has not been extrapolated to the whole industry, and that reports from different years should not be compared directly.Source: Australian Manufacturing source

③ Supply chain this week: joint-module capacity expansion moves from planning to production-line audits · hardware

This week two China A-share-listed transmission-component makers pushed their joint-module capacity a step forward. On September 29, a subsidiary of Wanliyang (Chinese transmission maker) passed a key customer's PPAP audit under the automotive VDA6.3 standard, so its first joint-module line is capable of mass production; phase one plans 100,000 units a year, and complete-robot customers are still at the engagement and prototype-testing stage. On September 28, Zhongda Leader (Chinese reducer maker) announced a private placement of up to RMB 530 million, of which RMB 250 million goes to an embodied-AI joint-module project with total investment of RMB 335 million, adding joint-module and planetary reducer lines, with a 24-month construction period. Overseas, Euronews, citing a report from Germany's Fraunhofer IPA, argues that German auto-parts makers are better suited to entering components such as joints, motors and dexterous hands, and that Schaeffler is already working with Rokae (Chinese robot arm maker) on humanoid joint modules. There is still a gap between planned capacity and actual orders: neither company has disclosed locked-in complete-robot orders. JPMorgan's report the same week favors scarce upstream components, on the condition that this capacity can win repeat orders.Source: 21st Century Business Herald; Jiemian News; Euronews source; Caiwen source