Today's Highlights

· 15 humanoids from Leju (Chinese humanoid robot maker) are loading parts on an auto-component line at Zhaofeng Electromechanical (Chinese auto-parts maker), with about 1,000 cumulative operating hours

· 291 Hai Robotics (Chinese warehouse-robot maker) HaiPick climbing robots have gone into Anta's (Chinese sportswear brand) Langfang warehouse, providing 120,000 storage slots

· Tesla's Robotaxi crash data shows 5 remote takeovers so far, and every one ended in a collision

· AgiBot's (Chinese humanoid robot startup) Lingxi X2 has served as a triage guide at Peking University People's Hospital for 66 working days, with 46,686 interactions

· REKI, a 2.4B VLA pretrained only on public data, is released with its code and all weights

· MolmoBot's action expert cut its conditioning layers from 36 to 3, saving up to 33% of inference compute with no change in success rate

· Fa'ao Robotics (Chinese collaborative-robot maker) closed C+ and D rounds in succession, totaling nearly RMB 1 billion

REKI: a 2.4B VLA pretrained only on public data, with code and weights fully released · vla

REKI releases everything at once: all the code from data cleaning through training, inference, evaluation and deployment, plus all pretrained and fine-tuned weights. Pretraining uses only public robot data, first-person human data and vision-language data. On LIBERO-Plus, RoboTwin Clean2Rand and RoboCasa365, only models about twice its size beat it. The action expert adds a few register tokens that remember the most recent frames; on a real robot, the policy swings back and forth between two poses about four times less often.

Juliette Marrie et al. (Kyutai) · arXiv 2610.22178 source

Taking apart three VLAs to trace causal pathways: cutting 90% of conditioning layers leaves success rate unchanged · vla

In π0.5, MolmoAct2 and MolmoBot, the causal influence of images and instructions on actions is concentrated in a narrow band of cross-attention layers in the middle-to-late part of the network, which the authors call the "action-conditioning interface." The representation of the target noun in early-to-middle layers determines which object gets grasped, and the spatial information in text tokens in middle-to-late layers determines where to move. Based on this, they reduced the conditioning input of MolmoBot's action expert from 36 layers to 3, cutting inference FLOPs by up to 33%; the success rate over 3,000 episodes is 46.7%, versus 46.9% for the default model. The principle also holds in real-robot deployment on π0.5.

Aiden Swann et al. · arXiv 2610.20750 source

Bench2Drive's comfort metric actually measures how long the car is stopped · benchmark

Bench2Drive is the mainstream benchmark for closed-loop evaluation of end-to-end driving. Its comfort metric is adapted from nuPlan, contains several implementation errors, and its segmentation scheme rewards standing still. Across 8 models, the metric's correlation with stationary time is as high as r=0.97. With a corrected version computed per distance driven, the correlation drops to 0.06; when models are compared at the same speed, the previously reported trade-off between driving score and comfort can no longer be detected. The authors also found two other problems: the default controller may replace the model's own steering with an angle pointing at the route target point, and the target points given at evaluation time may differ greatly from those used in training. The two issues affect UniAD and VAD very differently. The corrected code will be released separately and can recompute results from existing trajectory logs.

Thomas Schwarzfischer et al. · arXiv 2610.21741 source

DSP: a few minutes of human "play" data lets a dexterous hand collect its own interaction data · manipulation

The robot first behavior-clones a play policy from a few minutes of human play data, then runs it autonomously for several hours, and the interaction data it collects is used to train a world model. At inference time, the same policy supplies candidate action sequences for MPC. In real-robot experiments, the success rate is 69.6% higher than hindsight behavior cloning and on par with reinforcement learning, while using 490 times less interaction data. The whole pipeline needs no task demonstrations and no sim2real transfer. Jitendra Malik is among the authors.

Botao He et al. · arXiv 2610.22340 source

Spark-0: 1,000 hours of whole-hand tactile data, with touch as pretraining supervision · manipulation

Touch serves two purposes here. During pretraining, contact maps from tactile gloves supervise the visual tokens together with a frozen 3D-geometry teacher, and this can be removed at deployment; a separate tactile pathway encodes the force history of each region and passes it to the action expert. For this the team collected HumanTouch, a 1,000-hour dataset of whole-hand tactile human manipulation. With the same data and compute, keeping the tactile supervision raises the success rate on Ego Humanoid under vision-only deployment from 52.5% to 71.7%. The evaluation suite SparkArena is also released.

Tianxing Chen et al. · arXiv 2610.22342 source

Whole-Body Dexterity: clamping large objects with the arms and torso · locomotion

A Unitree (Chinese robot maker) G1 clamps a cardboard box under its arms and carries it from a shelf to a table, using its arms and torso. The team first generates a batch of feasible whole-body grasps, trains RL experts to reposition from any grasp to any grasp, and then distills them into a single policy that sees only proprioception and object geometry, transferring zero-shot to the real robot. The real-robot whole-body manipulation task completion rate is 90%, versus 18% for direct retargeting and 4% for a general whole-body controller.

Conor Mc Gartoll et al. · arXiv 2610.22184 source

VLA-Δ: treating changes in physical state as the training target · vla

HF↑36, the most upvoted paper in the community today. Most VLAs map observations directly to demonstrated actions; VLA-Δ instead uses the physical change between the current state and future states as explicit supervision. To do so it builds a large-scale pretraining set with state-transition annotations, with no added inference overhead. Results on LIBERO and RoboTwin are on par with SOTA, and in real-robot experiments it surpasses π0.5. The code is open-sourced.

Qi Lu et al. (ZJU-OmniAI) · arXiv 2610.21306 source

CarryOn: adding two outer loops to a frozen VLA · manipulation

The planning loop remembers user preferences and scene context; the execution loop predicts dangerous contact and blocks the corresponding actions, calling reusable recovery skills on failure. For a VLA that has already been safety-fine-tuned, the violation rate falls from 52.6% to 6.1%. After reusing reflection-improved skills, the recovery success rate rises from 13% to 80%. HF↑8.

Zixuan Wang et al. · arXiv 2610.21453 source

T2-z: inferring unseen environment parameters from a single interaction trajectory · world-model

Unobservable environmental factors such as friction and mass are carried by a learnable environment latent variable. At test time the world model stays frozen, and the latent variable for a new environment is optimized from just one interaction trajectory. Across 7 simulated settings and 4 real-robot tasks, the action success rate rises from 33.73% to 72.96%.

Chenyuan Zhou et al. · arXiv 2610.22364 source

Fa'ao Robotics | Series C+ and D | nearly RMB 1 billion | valuation undisclosed · industrial

The two rounds were led by Meituan Longzhu (venture arm of Meituan, Chinese delivery giant) and Meituan respectively, with Yongxin Capital, Shishan Phase II and Su Chuangtou participating; existing shareholders including Guolian Minsheng Capital, Suchan Investment and Qingliu Capital added to their stakes. The funds will go mainly to core-component capacity and to expanding production and R&D of the ART series of industrial-grade humanoid force-control arms. The arm has already signed cooperation deals for 2,000 units each with Boyin Hechuang and Jack Technology. Fa'ao just completed its A-share IPO coaching filing in China in September, and the company expects global order sales of 18,000 units in 2026, up 68% year on year, with Europe accounting for 46% of overseas sales.Source: Gasgoo source; 36Kr source

Mach Heart Technology (Li Auto's chip subsidiary) | first round | RMB 1.625 billion | post-money up to RMB 17.1 billion · hardware

Led by Hangzhou Jingtao, Lanchi Ventures and Huaye Tiancheng, with MiraclePlus (Chinese startup accelerator) participating. Hangzhou Jingtao and affiliated funds also received an option to subscribe for up to a further RMB 500 million; if fully exercised, the post-money valuation will be RMB 17.1 billion. Li Auto (Chinese EV maker) still holds 74.01%. The Mach M100 uses a 5nm automotive-grade process, delivers 1,280 TOPS per chip, and is already installed in Li Auto's new-generation production vehicles. The announcement states that the funds are especially to be used to expand into non-automotive fields.Source: Ifeng Auto source

XHEART (Xinxin Hangtu, Chinese automotive-chip startup) | strategic round | amount undisclosed | about RMB 5 billion cumulative · hardware ⚠️ Rumor-based

36Kr, citing multiple industry sources, says that in this round Renault Group invested for the first time, Chery (Chinese automaker) added to its stake, and the memory maker CXMT (Changxin Technology, Chinese memory chipmaker) also took part. The flagship chip X7 targets L2++ urban driver assistance and L3/L4, and was developed in close collaboration with Momenta (Chinese autonomous-driving company) from the design stage. The X7's first vehicle is the SAIC Volkswagen ID.ERA 9X, followed by MG, Roewe, Cadillac and some Chery models. With memory prices rising this year, 36Kr believes CXMT's stake may give it an edge in memory pricing and supply.Source: 36Kr source

CMR Surgical (UK) | strategic investment from Stryker | about $400 million | with acquisition option · adjacent

Through this investment, Stryker enters soft-tissue surgical robotics and also obtains an option to acquire CMR outright in the future. The two sides also announced a strategic alliance to jointly advance robot-assisted surgery.Source: The Robot Report source; MedTech Dive

Youai Zhihe (Chinese mobile-robot maker) | Hong Kong IPO | CSRC filing obtained | up to 49.4778 million shares planned · industrial

On October 16, the China Securities Regulatory Commission issued a filing notice for overseas listing and "full circulation," under which about 244 million domestic shares held by 37 shareholders will be converted into overseas-listed shares. The company is applying as a specialist technology company under Chapter 18C of the Hong Kong Stock Exchange listing rules, with CICC (China International Capital Corporation) as sole sponsor. Revenue for 2023 to 2025 was RMB 108 million, RMB 255 million and RMB 340 million respectively, with gross margin rising from 26.1% to 36.5%. Its main business is mobile manipulation robots, used at semiconductor, lithium-battery and energy-chemical sites.Source: Sina Finance source; Stockstar source

Sichuan Embodied Humanoid Robot (EIR) | Pre-A round | over RMB 100 million | valuation undisclosed · humanoid

The company was incubated by the robotics research center of the University of Electronic Science and Technology of China and founded in September 2024; a new financing round has been launched in parallel, with investors undisclosed. The funds will go toward iterating the emotionally interactive humanoid "Aiqiu," a production base with capacity in the ten-thousand-unit range, and a sales system. Aiqiu went on sale in August, starting at RMB 98,000; the complete-machine production base is expected to start operating by year-end, beginning at annual output in the thousands of units.Source: Cover News source

Noble Machines (US) | strategic investment from Shimizu Corporation | amount undisclosed · humanoid

The Japanese construction company Shimizu Corporation has taken a stake and will co-develop humanoid robots for construction sites with the company. By the end of the year they will first run virtual simulations using BIM/CIM data, and next spring move to testing on real construction sites. The initial tasks are carrying materials, walking around the site and patrolling. Noble Machines was founded in May 2024 in Sunnyvale, California, its CEO is Wei Ding, and its products are the MOBY series of construction humanoids.Source: AI Insider source

Panhu Intelligence (Hangzhou) | angel round | tens of millions of RMB | valuation undisclosed · world-model

Invested in by Puhua Capital and Hengze Capital. The company was founded this year and is developing TigerMind, a world model that fuses vision, force and touch. Founder Hou Dibo is a professor at the College of Control Science and Engineering at Zhejiang University; co-founder Lin Feng was formerly a vice president at DingTalk (Alibaba's workplace app) and CTO of Alibaba's B2B search and advertising.Source: Zhidx source

15 Leju humanoids load parts on Zhaofeng Electromechanical's production line, with about 1,000 cumulative operating hours · humanoid

15 Leju humanoid robots pick, transfer and load parts on Zhaofeng Electromechanical's auto-component production line, covering 15 workstations and more than 10 types of parts. Leju says a single loading cycle takes under 40 seconds with a success rate of about 98%, and that cumulative operation of about 1,000 hours has helped produce more than 60,000 parts. Operation data is sent back in sync for quality classification and model iteration. The two sides also donated the real-robot loading dataset to the OpenLET community of the OpenAtom Foundation. The project is part of the on-site training campaign for humanoid robots and embodied intelligence run by China's Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission of the State Council, which requires key products to complete application verification and routine deployment in representative scenarios by the end of 2026. The report also states that whether this solution can be replicated in other scenarios remains to be verified.Source: Sohu source; Pandaily source

291 Hai Robotics HaiPick robots go into Anta's Langfang warehouse · industrial

Anta's Langfang regional warehouse covers more than 6,400 square meters and has deployed 291 Hai Robotics HaiPick robots, with 78 charging stations, 5 inbound workstations and 32 outbound workstations. Storage height has been raised to 8.7 meters, giving 120,000 storage slots, and Hai Robotics says storage density has increased 3-fold. According to figures published by Hai Robotics, outbound throughput is 4,417 cartons per hour, inbound is 1,063 cartons per hour, and daily capacity exceeds 210,000 items. At the end of August, Hai Robotics said this product had signed more than 10,000 units globally in cumulative contracts, and Anta was already on the customer list at that time.Source: AI Cloud News source; EE Times China source

AgiBot Lingxi X2 serves 66 working days as a triage guide at Peking University People's Hospital · humanoid

Operating statistics from Peking University People's Hospital show that, as of September 30, Lingxi X2 had worked 66 working days at the outpatient navigation desk, with 46,686 cumulative interactions, averaging about 707 per day and peaking at 1,420 in a single day. The robot is connected to the AI medical assistant in Ant Group's Afu app and to the hospital's private knowledge base, answering questions on department guidance, visit procedures, medical-insurance policy and in-hospital routes. These figures count only the number of interactions; they give no answer accuracy and do not say how many manual inquiries were offloaded.Source: China.com Finance source; PEdaily source

Tesla Robotaxi: 5 remote takeovers, 5 crashes · autonomy

In Tesla's latest crash reports filed with NHTSA, two more Robotaxi collisions occurred while a remote operator was in control. In Dallas in August, after being taken over, a vehicle scraped an SUV at about 0.4 mph; in Austin the same month, one hit a trash bin at about 1 mph. Neither vehicle had a passenger aboard. Electrek counts 5 remote takeovers in the crash data, all ending in collisions; the earlier three hit a tree stump, an iron fence and a construction barrier. Tesla does not disclose how often it intervenes remotely, so no rate can be calculated. In a letter to Senator Markey in February this year, Waymo wrote that it "does not use remote driving in which a human performs the dynamic driving task." Its only exception is moving stuck vehicles at 2 mph with a fixed steering angle, and it said the tool has never been used outside of training. Tesla's cumulative Robotaxi crash reports now number 36, up from 22 in July.Source: Electrek source

7 driverless rigid-frame mining trucks at the Baiyinhua coal mine; Zhitong Finance says the system comes from Yikong Zhijia · autonomy

Seven TR100A rigid-frame mining trucks at an open-pit coal mine in Baiyinhua, Xilingol League, Inner Mongolia, have been operating continuously with no safety operator on board. Zhitong Finance says the system was supplied by Yikong Zhijia (Chinese autonomous mining-truck company), and that this is its first routinely operating driverless rigid-frame truck project in China; Yikong has cumulatively deployed more than 3,900 driverless mining trucks across over 40 mines, and on the day the news came out its Hong Kong-listed shares rose more than 10% intraday.Source: Stockstar source

DeepMind builds a robotics hardware lab in Cambridge, led by a former Boston Dynamics CTO · embodied

Aaron Saunders announced on LinkedIn that Google DeepMind is building a robotics hardware lab in Cambridge, Massachusetts. The new team aims to put Gemini models into machines that walk and grasp. According to The Boston Globe, Saunders revealed the lab could open in 2027 but gave no specific location or size; it is hiring machinists, engineers and researchers. Saunders worked at Boston Dynamics for 22 years and served as CTO from 2021, in charge of Atlas and Spot; he joined DeepMind in 2025 and is now vice president of embodied systems and robotics. At the end of 2025, Hassabis described DeepMind's robotics approach as the "Android play," aiming to make Gemini a general-purpose system that can run on many kinds of robot bodies.Source: Hoodline source

Waymo picks Toronto as its first Canadian city, targeting passenger service in 2028 · autonomy ⚠️ Plan-based

Waymo's vehicles began hitting the road in Toronto in recent days, first driven manually to map the area and validate software, with the goal of welcoming its first public passengers in 2028. Electrek points out that Waymo's announcement mentions no permits, while Ontario's autonomous-driving pilot regulation expires on October 13, 2027, and the provincial government has not said what rules will replace it.Source: Electrek source; Reuters source

Volvo expands its Nvidia partnership to build door-to-door autonomous driving · autonomy

Volvo Cars announced on October 15 that future models will connect its own safety systems to Nvidia's Hyperion sensing and compute architecture, Halos OS and Alpamayo driving software, covering highways, busy city streets and parking scenarios. Alpamayo is a VLA model that outputs a trajectory along with a causal reasoning chain, which can be audited during safety validation. CEO Håkan Samuelsson said that for nearly a century Volvo has been synonymous with automotive safety. Nvidia also expanded its autonomous-driving collaboration with Mercedes-Benz around the same time.Source: Unite.AI source; AI Insider source

UBTech (Chinese humanoid robot maker) says Thinker-WM ranks first on three world-model benchmarks · world-model ⚠️ Vendor-reported

WorldOlympiad composite score 0.740; EWMBench total score 4.66, surpassing Qwen-RobotWorld and Sora2; leading overall across the 12 dimensions of WorldArena1.0 Track-1. Thinker-WM uses a Unified Dynamic Transformer to encode RGB video, depth and robot actions into one representation space. The training corpus is about 30,000 hours, of which about 20,000 hours are first-person human video and about 10,000 hours are real-robot data. These rankings are so far published only by UBTech.Source: Guandian source

Two class actions in Chicago: delivery robots accused of illegally collecting biometric data · adjacent

Plaintiffs allege that the delivery robots of Coco and Serve Robotics collect pedestrians' face and hand data through cameras, violating the Illinois Biometric Information Privacy Act. A Coco spokesperson said the suit "fundamentally misunderstands our technology," and that the company does not collect biometric data in Chicago. Serve says its system is not designed to identify individuals and that it will contest the suit. A petition calling for a pause to the sidewalk-robot pilot has more than 4,600 resident signatures. Chicago's pilot expires in May 2027, unless the city council ends it early or decides to extend it.Source: CBS News source

Faraday Future humanoid FF Nova priced at $2,490 · humanoid

The nine robots launched at an event in September were priced between $9,990 and $137,900; this time the FF Nova is priced at $2,490, and the education model Nova Mini, released at the same time, sells for $299. Nova is sold in three phases: the first phase is limited to 100 units for developers and educational institutions, and Faraday Future says the first 20 units are delivered immediately; the consumer version is slated for the third phase.Source: Guandian source; Unite.AI source

Fincantieri plans to invest nearly $100 million by 2030 to put robots in its shipyards · industrial ⚠️ Plan-based

The investment of about $100 million falls under BOOST², a project the Italian shipbuilder launched three years ago. The project already covers 13 shipyards and more than 450 employees, including Marinette Marine in Wisconsin, US. The next phase shifts from technical trials to replicable industrial deployment, with collaborative robots used to help workers with complex production tasks. CEO Pierroberto Folgiero said the challenge now is to "accelerate scaled deployment and adoption."Source: AI Insider source

Minneapolis veto override fails; Robotaxis need not carry a person · autonomy

The city council voted 7 to 5 on overriding Mayor Jacob Frey's veto, 2 votes short of the 9 required. The ordinance requiring Waymo, Tesla and Zoox to obtain city permits and to keep a person in the driver's seat therefore cannot take effect.Source: KSTP source; Minnesota Reformer source