Today's Highlights

· Musk says Tesla has already beaten Waymo in Austin; on a nationwide basis, Waymo's weekly driverless miles are still about 7x Tesla's

· ANYbotics' explosion-proof quadruped ANYmal X begins deliveries to Equinor and DSM-Firmenich

· Physical AI investment in China fell 40.1% in September alone, and a state-owned Chinese investor says it has stopped looking at new projects since June

· UBTech (Chinese humanoid robot maker) sold 921 full-size humanoids in the first half, 10% fewer than in the second half of last year

· ZAP was trained with no real-robot data and reaches a 96.7% real-robot success rate, on par with GPT-6 Astra

· Imperial College spinout Machenta raises a $40 million seed round, the largest ever for a UK robotics company

RynnVLA-Latent: pretraining a VLA on latent actions from 30,000 hours of videovla

HF 9↑. Robot demonstrations with action labels are scarce, so RynnVLA-Latent instead extracts continuous 3D latent actions (an action representation that does not depend on a robot's native control signals) from the visual changes in human, robot, and simulation videos. It extracts 745 million latent-action descriptions from 30,100 hours of video, then uses that data to pretrain a 4B policy. The policy reaches a 98.8% success rate on LIBERO; moved to LIBERO-Plus without targeted adaptation, it still achieves 84.7%, and it scores 88.6% on the dual-arm benchmark RoboTwin 2.0.

Haoyu Zhao et al. · arXiv 2610.34356 source

ZAP: a small model with zero real-robot data matches GPT-6 Astravla

The team took an off-the-shelf 27B vision-language model and ran SFT and GRPO on simple simulated tasks only, teaching it to control a robot through a tool interface. No real-robot data was used at any point in training, yet real-robot task success reaches 96.7%, on par with GPT-6 Astra under the same interface, while MolmoAct2 manages only 33.3%. It is 3 times faster than GPT-6 Astra at about one-seventeenth of the cost. On a single-arm Franka it had never seen, it scores 55.9% on RoboLab-120, 4.9 percentage points above the best VLA/WAM baseline.

Benjamin Pikus et al. · arXiv 2610.35817 source

PhysLadder: GPT-6 Astra drops to zero when it must "switch contacts while applying force"benchmark

16 household tasks are split into three levels, with simulation and real-robot environments aligned one to one. Zero-shot GPT-6 Astra declines level by level on the real robot: L1, choosing contact points and aligning geometry, 80.0% success; L2, regulating force during sustained contact, 47.6%; L3, controlling force while switching contacts, 0%. For 10 of the 16 tasks, the number of successes is exactly the same in simulation and on the real robot, which leads the authors to argue that a well-matched simulation can replace part of real-robot testing.

Haibao Yu et al. · arXiv 2610.34527 source

EBench: GPT-6 Astra controlling the robot itself, or calling a VLAbenchmark

This technical report from InternRobotics (Shanghai AI Laboratory's robotics team) compares two ways of using the model. When GPT-6-Astra outputs waypoints directly, it leads on short-horizon tasks but trails post-trained policies on long-horizon, tabletop, and high-precision contact tasks. Switched to calling a VLA or WAM on demand, its overall success rate catches up to that of the strongest post-trained policy, and token cost also drops sharply. The failures come mainly from badly timed calls, and from knocking away an already-grasped object when it takes back control.

Yifei Yao et al. · arXiv 2610.35832 source

EmbodiedLab: getting agents to run physical experiments hands-onbenchmark

HF 9↑. The benchmark covers 30 quantitative experiments across five disciplines and 14 types of instrument platforms, with system parameters re-sampled each run so results cannot be memorized in advance. Six frontier models operate simulated labs inside their own coding agents; even the strongest succeeds less than half the time, and most failures occur at the step of operating the instruments. The benchmark and evaluation code are open-sourced.

Hongcheng Gao et al. · arXiv 2610.35385 source

Zero-HLM: scan a scene with a phone, and a humanoid trained on simulated data beats one trained on real datalocomotion

Scenes captured on a phone are reconstructed and paired with simulated demonstrations, generating 4,800 physics-verified loco-manipulation trajectories. On three tasks on the Unitree (Chinese robot maker) G1, GR00T N1.7 trained only on this data reaches a 66.7% all-task success rate, versus 46.7% when trained on real-robot data. In a generalization setting, the real-data baseline reaches only 23%, rising to 69% after the simulated data is added.

Dianyi Yang et al. · arXiv 2610.35075 source

FLORE: getting humanoid hands to work by forcelocomotion

For tasks like wiping, pushing, and hugging, trajectories alone are not enough. FLORE collects demonstrations with two UMI-style handheld grippers with force sensing and wearable trackers, with no robot present. The policy outputs both actions and a force command for each hand, which a compliant whole-body controller then executes. On three contact-rich tasks the success rate is 80%, versus only 26.7% for a position-only policy trained on the same data.

Sihang Li et al. · arXiv 2610.35862 source

VLA-HWBench: edge VLA latency is determined mainly by how actions are generatedbenchmark

The paper benchmarks 18 open-source VLAs on Jetson AGX Thor and AMD Ryzen AI Max+, timing inference in seven segments. On Jetson, latency differs by up to 82x between models, and parameter count explains at most 15% of the variance. Models that generate a long action chunk at once with few-step flow matching keep up with the control rate; models that decode token by token do not. Operator fusion plus 4-bit weights speeds up π0.5 and GR00T N1.7 by 2.6 to 2.9 times, but ready-made implementations currently exist only for these two models and only on Jetson.

Borui Xu et al. · arXiv 2610.34717 source

What world modeling actually contributes to action learningworld-model

In FastWAM and OpenWAM, the authors remove only the future-video-prediction loss and hold everything else fixed. Success on seen tasks is higher by 24.38 and 72.91 percentage points respectively, but no stable advantage appears once environment randomization is changed. Adding a lightweight prediction head used only during training to π0.5 lifts success from 82.94% to 89.19%, with no need for a large video DiT.

Hanwen Cui et al. · arXiv 2610.34593 source

Machenta ｜ Seed ｜ $40 million ｜ Valuation undisclosedworld-model

Co-led by Creandum and Union Square Ventures, in what Imperial College calls the largest seed round for a UK robotics startup. Co-founder and CEO Edward Johns, a professor of robot learning at Imperial College, says: "Intelligence isn't about what you already know how to do, but how quickly you learn something new." The company's models are pretrained entirely on simulated data, "without a single real-world pixel," according to Johns, and can perform new tasks such as box packing and tidying a table after seeing one demonstration. The company has only 11 people, is focused on research for now, and has no near-term commercialization plans; Johns also admits that high-precision tasks and complex motions are still not handled well. Sources: Imperial College London source; Business Insider source

Bedrock Robotics ｜ New round ｜ $600 million ｜ Valuation $4.1 billionautonomy

Axios reported in mid-October that it was raising; the round has now closed at a valuation of $4.1 billion. The company makes autonomous excavators, targeting construction sites where heavy-equipment operators are aging and hard to hire. Sources: The Tech Buzz source; Ebrun (Chinese e-commerce news site) source

FF robotics business ｜ Merger into FFR completed ｜ Valuation about $200 millionhumanoid ⚠️ Vendor claim

Following the definitive agreement signed in mid-October, Faraday Future's robotics business and related assets were transferred to Nasdaq-listed FFR on October 23, in exchange for FFR common stock plus non-voting convertible preferred stock. Jia Yueting (Faraday Future founder) becomes FFR CEO and unveiled FF OmniMind, a unified robot brain. The company says October robot sales will exceed 300 units, on the same basis as before, including humanoid and biomimetic robots. Sources: Business Wire source; Huxiu (Chinese tech media) source

Weifa Electronics ｜ Series C ｜ RMB 100 million-plus ｜ Valuation undisclosedhardware

Guokexinneng and Huatai Zijin are new investors, while Jianfa Xinxing Investment, Yida Capital, and Sugaoxin Venture Capital added to their holdings. The Changzhou-based company's eddy-current sensors are already in mass production on more than ten new-energy vehicle models, and it is now applying the same technology and automotive-grade system to humanoid joint encoders, launching three series (single-turn absolute, mechanical-gear multi-turn, and same-side dual-encoder) now being tested with leading customers. A single humanoid robot often needs dozens of encoders. Source: PEdaily (Chinese venture-capital news site) source

Jinkong Technology ｜ Angel ｜ RMB tens of millions ｜ Valuation undisclosedadjacent ⚠️ Plan-based claim

Led by Hony VGC (venture arm of Chinese PE firm Hony Capital), with participation from Shanghai Jiao Tong University's Hanyuan Asset Management, Guoke Green Fund, and Lingqiao Intelligence. Founded in April, the company builds space robots; its founder is Shanghai Jiao Tong University professor He Jun. Under the plan, the first batch of core components, including controllers and servo drives, will go into orbit for testing aboard a computing satellite at the end of this year, and in mid-2027 "Jinkong-1" will demonstrate autonomous tool changing in orbit. Source: EE Times China source

VLAI Future Power ｜ Series A ｜ Amount undisclosedhardware

Tianyancha records show the Nanshan Strategic Emerging Industry Investment Fund and the Futian Guidance Fund participated. The team comes from HKUST (Guangzhou) and makes integrated joint motors for humanoid, quadruped, and wheel-legged robots. Source: Stockstar (Chinese financial news site) source

Musk says Austin "was never the real contest"; nationwide, Waymo's miles are still about 7x Tesla'sautonomy

In the week ending October 18, Tesla's unsupervised miles in Austin were 438,000 miles, up 30% week over week, exceeding Waymo's Q2 local average of about 406,000 miles per week. Musk replied on X: "This was never a real contest." Electrek points out that this compares Tesla's latest week with Waymo's quarterly average from four months earlier. On a nationwide basis, Waymo ran 3.9 million miles per week in Q2, while Tesla's best week was 558,000 miles. 79% of Tesla's driverless miles are concentrated in Austin, while Houston, Tampa, Orlando, and Miami together total about 25,000 miles per week. The direct cause of the jump in Austin mileage is the Cybercab: 463 vehicles are now registered on the Texas DMV list, 337 of them added in the last 30 days. Tesla's Lars Moravy said during a Munro Live factory tour: "The line is running so fast that we don't actually run it every day; production capacity now exceeds the pace of deployment." He explained that Cybercab final assembly has only 25 stations and about 78% of the processes are automated. Sources: Electrek source; Drive Tesla Canada source

ANYbotics' explosion-proof quadruped ANYmal X begins deliveriesindustrial

The first customers include Equinor and DSM-Firmenich, with delivery to hazardous areas of oil, gas, and chemical plants where ordinary robots are not permitted. Since 2023, BP, Equinor, Shell, PETRONAS, Petrobras, BASF, and SLB have run hundreds of missions onshore and offshore through an early-access program. The robot carries thermal imaging, vision, acoustic, and gas sensors, logs readings equipment by equipment, and connects to the plant's existing maintenance software. Co-founder and CEO Péter Fankhauser says: "Many experts believed legged robots could never be made safe in explosive environments." The company calls it the world's first legged inspection robot for explosive environments; in September, Ruigan Robotics (Chinese robotics company) obtained China's first dual gas-and-dust explosion-proof certification for a wheeled-legged robot. Source: Unite.AI source

Topstar signs about 10,000 industrial robot orders in the first three quartersindustrial

Orders of about 10,000 units and shipments of about 8,600 units, with orders for in-house articulated robots up about 34.44% year on year. Revenue from industrial robots and automation application systems was RMB 1.028 billion, up 88.83% year on year, mainly from leading 3C customers; net profit attributable to shareholders was RMB 146 million, up 196.38%. CNC machine tool orders were about 390 units, and the company attributes the growth to rising demand for machining humanoid robot components. Topstar (Chinese industrial automation maker) Source: OFweek Robotics source

Waymo enters Las Vegas airportautonomy

Since October 21, Waymo passengers have been picked up and dropped off at the dedicated zones at Harry Reid International Airport's T1 and T3, initially open only to existing local users, who can travel anywhere within the 24-square-mile service area. This is the sixth major airport Waymo has joined. Waymo has not released trip volumes, the size of its airport fleet, or a date for full opening. Zoox began operating at the same airport in September. Source: Transportation and Logistics International source

Fukuoka Airport's L4 cargo tractor goes live, starting with 1 vehicleautonomy

Japan Airlines and Fukuoka International Airport put an L4 autonomous tractor into service on a cargo route on October 26. The route is about 2.5 km in total, of which 1.7 km is driven by the system. There is currently just 1 vehicle, with plans to increase to 4 by year-end. After Narita and Haneda, this is the third airport in Japan to do so. Source: Digital Today source

Cainiao delivers two apparel climbing-robot warehouses ahead of Singles' Dayindustrial

One is in Tianjin, a returns warehouse for a leading apparel brand; the other is in Suzhou, for an international sportswear brand. Cainiao (Alibaba's logistics arm) launched this climbing robot in April and has already delivered projects in Dongguan, Guangzhou, Hong Kong, Spain, and the Netherlands; its first climbing warehouse in Vietnam is expected to be delivered at the end of December. Cainiao has not disclosed how many robots each warehouse has. Source: Cainiao source

Edeka meat distribution center adds two Knapp picking robotsindustrial

Two Pick-it-Easy Robots pick fresh meat around the clock, each at about 450 picks per hour, scanning and weighing along the way. This Nordfrische center east of Hamburg supplies about 2,500 products a day to roughly 650 stores, and the whole system handles up to 40,000 totes a day at +2°C. Source: Robotics & Automation News source

Physical AI investment in China fell 40.1% in September alone, while global robotics venture funding is still risingadjacent

According to Securities Times (Chinese financial newspaper) and data firm NE Times, physical AI investment in China fell from RMB 19.4 billion in August to RMB 11.6 billion in September, and the top ten projects' share fell from 67.7% to 59.4%. Seed, angel, and Pre-A rounds instead rose from 19 to 24. Separately, according to Yicai (Chinese business media), the third quarter totaled RMB 33.335 billion, 28.7% below the first-half quarterly average. An AI investment manager at a state-owned investment institution says they have effectively stopped reviewing physical AI projects since June. Zhang Zhikuan, a director at Co-Stone Capital, an AgiBot investor, says many companies already carry high valuations and follow-on investment needs to be more cautious. The secondary market is also loosening: an article republished by PEdaily, citing Rongzhong Finance, reports that an investor who bought into an embodied-AI company at a RMB 3 billion post-money valuation earlier this year now wants to sell the shares at RMB 12 billion, about 45% below the company's publicly stated valuation of RMB 22 billion, and still has not found a buyer. This is only an asking price, not a completed transaction. Crunchbase data points the other way: global robotics startup venture funding in the third quarter was $9.9 billion, more than double the $4.3 billion a year earlier, with a cumulative $13.2 billion in China and $12.1 billion in the US this year. The two sets of figures differ in scope and period: one is monthly and for China, the other quarterly and global. Sources: Seoul Economic Daily source; Crunchbase News source; PEdaily source

UBTech's full-size humanoid sales fall 10% half over half, with average unit price down to about RMB 640,000humanoid

In the first half of 2026 it sold 921 full-size humanoids. Working back from 1,079 units for full-year 2025, the second half of last year was 1,034 units, from which Shizhi Fengyun (Chinese financial media) calculates a half-over-half decline of about 10.9% in units and about 24.6% in revenue. Dividing revenue by units, the average unit price fell from about RMB 849,000 in the first half of last year to about RMB 641,000; this average includes solution revenue. Management raised this year's Walker S series delivery target from 2,000–3,000 units to 5,000 units and wrote it into the equity incentive conditions in September. The U1 reportedly has orders of more than 13,000 units, but management expects to deliver only 1,500 to 2,000 within the year. Inventory rose from RMB 577 million to RMB 985 million, and a single customer contributed 24% of revenue. Source: Shizhi Fengyun source

AgiBot releases embodied agent framework Genie Master; G2 adapts on the job at a convenience storeembodied ⚠️ Vendor claim

While hanging goods, the head camera could not clearly see hooks blocked by labels; the agent reviewed its failure records and changed its approach: first approach and locate with the left wrist camera, then have the right hand hang the goods. AgiBot says the framework uses an execution protocol, EEP, to tie together agent judgment, robot execution, and result feedback; part of the experience is written to long-term memory and part is fed back to train the action policy. The company claims this is the first validation of "embodied recursive self-evolution" in a real business setting. On November 2, six Genie G2 robots will go on duty at a robot convenience store for a 24-hour continuous livestream. Source: IPO Zaozhidao (Chinese tech and venture media) source

KDDI sets up Elyza Robotics, plans to invest ¥30 billion in physical AIworld-model ⚠️ Plan-based claim

The new company began operating on October 1, with a target of deploying more than 1,000 robots in the field by fiscal 2030 and expanding the group's physical AI team to more than 300 people. It will develop VLM and VLA foundation models, and this year will first run proofs of concept with industrial robots, quadrupeds, and humanoids from Japanese and overseas makers, including restocking and inventory management at Lawson stores. Source: AI Insider source

NEURA launches 4NE1 Gen 4, putting compute in every armhumanoid

Each limb has a Qualcomm Dragonwing processor at the joints, so collision reflexes, grip adjustment, and balance correction are handled locally, while planning is left to the foundation model Neuron AI. The outer shell is switched to fabric, which the company says is about 20% lighter than Gen 3; this figure has not been independently measured. The launch took place at TUM RoboGym, which opened on October 22; this 2,300-square-meter training facility was built jointly by NEURA and the Technical University of Munich, with a total investment of about €17 million, of which NEURA contributed about €11 million. Source: Tech Times source

Samsung bets on five-fingered hands as Lee Jae-yong visits the RX divisionhumanoid

In July Samsung Electronics set up the RX division, which reports directly to the CEO, and is developing a humanoid robot with five tendon-driven fingers per hand; it has also recruited someone who previously led Boston Dynamics' robot strategy. Korean media report that Chairman Lee Jae-yong visited the Seoul Seocho R&D campus on the fourth anniversary of taking office and watched a demonstration of grasping with all ten fingers on both hands. Hyundai's Atlas has four fingers per hand, trading fewer fingers for simpler actuation, manufacturing, and maintenance. Sources: Hankyoreh source; Chosun Ilbo source

Baidu Apollo Go and Freenow show the RT6 in London, with commercial operation in 2027autonomy

The two began testing with safety drivers in London this summer, plan to remove the safety drivers next year, and will begin commercial operation in 2027, deploying only a few dozen vehicles over the next one to two years. Danny O'Gorman, general manager of Freenow UK, told Reuters: "The vehicles are at the stage of learning the rules of the road, much like an ordinary person studying for a driving test." Wayve and Uber launched a service with safety drivers in London in September, and Waymo plans to launch a fully driverless service there by year-end. Source: Gasgoo (Chinese automotive news site) source

IDC: global quadruped shipments exceed 49,000 units in the first half, with industrial-grade models taking two-thirds of revenueindustrial

Shipments rose 92.5% year on year and market size exceeds $400 million. Home and research models account for about 60% of shipments but less than one-third of revenue. Industrial-grade applications exceed $270 million, up 125.1%, with Deep Robotics (Chinese quadruped maker) ranked first in this segment. Source: Kuai Keji (Chinese tech news site) source