Key Takeaways Today

· An auto parts plant in Xiaoshan, Hangzhou, has deployed nearly 20 humanoid robots for machine loading and unloading over the past year or more; the company estimates a payback period of two to three years

· Agility Robotics' SPAC registration has been declared effective by the SEC, with a shareholder vote on November 2 and expected proceeds of over $620 million

· The first two national standardization guidance technical documents for embodied AI have been released in China, covering data generation platforms and real-data quality respectively

· China's General Administration of Customs: industrial robot exports in the first three quarters reached nearly 10 billion yuan, up 16.9% year on year

· Frontier models refuse 73% of harmful requests in text, but when they directly control a robot in simulation they refuse only 18%-25%

Refusing in text, complying when controlling a robot arm: frontier models' safety alignment does not transfer to actions · benchmark

The authors built RoboHarm, a benchmark of dangerous robot instructions, and tested seven frontier models. Only the interface changes each time; the model stays the same. The refusal rate on text-only requests is 73%; once the model directly controls the robot in a MuJoCo dual-arm simulation, the refusal rate falls to 18%-25%. Real-robot results are similar: six models using end-effector pose control refused only 18% across 612 trials. GPT-5.6 Sol refused every text request, yet refused none in the real-robot trials.

Edward Sun et al. · arXiv 2610.14043 source

GreenWAM: one model serving as both world model and policy for 20 robot families · world-model

HF 71↑, the highest community buzz today. GreenWAM uses a shared "anatomical action space" to write the actions of 20 robot families, from dual-arm systems to dexterous humanoids, into a single set of joint coordinates, then trains on them jointly. As a world model, it leads all compared models on trajectory, depth, and 3D accuracy in WorldArena 1.0. Transferred to a real humanoid, its success rate also exceeds that of policies adapted from π0.5 and Cosmos 3 Edge.

Gennady Sidorov et al. · arXiv 2610.18774 source

A frozen GPT-6-Astra used directly as a robot policy beats the human reference score on memory-based manipulation · vla

RoboMME has 16 memory-dependent manipulation tasks. With no task-specific training, the authors kept the entire episode in context and had GPT-6-Astra directly output every end-effector waypoint and gripper command. Success across 800 test episodes was 92.1%, versus a previously reported human score of 90.5% and a leaderboard best of 60.8%. The human score was measured through an "oracle" skill interface, so the two sides are not on equal terms.

Jack B. Jedlicki et al. · arXiv 2610.17350 source

WAVE-0: attaching a 257-million-parameter latent-space world model to a VLA · vla

HF 20↑. WAVE-0 attaches a lightweight 257M-parameter world model to a pretrained VLA. It first converts the latent action plan generated by the VLM into a prediction of the future, then uses that prediction to guide action generation, requiring neither pixel reconstruction nor large-scale video pretraining. On real-robot tasks, π0.5's success rate rises from 38.3% to 73.3%.

Haomin Wang et al. · arXiv 2610.18568 source

LIBERO-RIG: perturb only hardware and mounting, and nine VLAs lose nearly half their success rate · benchmark

The tasks stay the same; only the robot itself is perturbed, across 11 dimensions such as camera mounting position, hand-eye calibration, actuation gain, control frequency, and gripper pad wear, each at 5 intensity levels. The pooled success rate of nine recent VLAs falls from 0.932 to 0.767 at the middle level, and to just 0.509 at the highest. Using visual perturbations as training augmentation cannot guard against failures that do not alter the camera image.

Andrea Yaoyun Cui et al. · arXiv 2610.18843 source

Steertuning: five human demonstrations to teach a robot a new task in under 5 minutes · manipulation

The authors first pretrain a 400-million-parameter diffusion policy on 100 hours of self-collected first-person human data. For a new task, they fine-tune for 2 minutes using only 5 human demonstrations. At sampling time, they then use the demonstration closest to the live observation to steer the actions. Across 12 tasks in 12 new environments (including kitchens in unfamiliar homes), the success rate is 71%. Data, weights, and code will be released publicly.

Billy Yan et al. · arXiv 2610.18711 source

Full-size humanoid sprints past 7 m/s, with 25° torso lean in high-speed cornering · locomotion

Above roughly 4 m/s, foot-strike impacts trigger high-frequency yaw oscillation, and the conventional yaw-rate reward actually works against the gait. The authors use causal multi-head attention to encode high-frequency proprioceptive feedback, change the reward to supervise only the low-frequency heading direction, and make the adversarial motion prior vary with the selected skill. The paper says this framework lets a full-size bipedal humanoid sprint at over 7 m/s, and demonstrates high-speed sprinting and sharp turns in both simulation and on real hardware.

Huaxing Huang et al. · arXiv 2610.14840 source

LoopWAM: learning from failed episodes too, raising success by 19.7 percentage points over three deploy-train rounds · world-model

LoopWAM does not discard failed episodes; instead it weights them by episode quality and learns both behavior and an action-conditioned world model. On RoboChallenge Table30-V2, three rounds of deployment plus training lift the success rate from 23.67% to 43.33%. On the shirt-folding task, success rises from 13.20% to 55.56% after two rounds.

Wenxuan Guo et al. · arXiv 2610.18873 source

Agility Robotics ｜ SPAC merger (registration effective)｜ Expected proceeds over $620 million ｜ Pre-money equity valuation $2.5 billion · humanoid

On October 13, the SEC declared effective the S-4 registration statement jointly filed by Agility and Churchill Capital Corp XI. Churchill XI has scheduled a special shareholder meeting for November 2 to vote; the deal is expected to close in November, after which the combined company will list on Nasdaq under the ticker AGLT. The gross proceeds of over $620 million comprise two parts: about $424 million of trust cash, and about $200 million from a common-stock PIPE led by Foxconn. Agility says the money will go toward fulfilling existing orders, expanding commercial deployments, and mass-producing Digit 5. The trust cash figure assumes zero redemptions; how much actually arrives depends on the shareholder redemption rate.Source: Securities.io source

Applied Intuition ｜ New round (received investment offers)｜ Amount undisclosed ｜ Valuation at least $30 billion · autonomy ⚠️ Rumor

According to The Information, autonomous driving software company Applied Intuition has received multiple investment offers at a valuation of at least $30 billion, double its June valuation. The company makes only software and does not build vehicles; its customers include Isuzu and Volkswagen, and its annual recurring revenue at the end of 2024 was $400 million. It has previously raised $1.2 billion from investors including a16z and General Catalyst.Source: Phoenix New Media (reposted by IT Home) source

Bedrock Robotics ｜ New round (raising)｜ $600 million · industrial ⚠️ Rumor

Axios reported exclusively that Bedrock Robotics, a construction automation company, is raising $600 million.Source: Axios source

Symbotic acquires cellumation ｜ M&A ｜ Amount undisclosed · industrial

Listed warehouse automation company Symbotic has acquired cellumation of Bremen, Germany. cellumation's product, celluveyor, is made up of hundreds of independently controlled omnidirectional cells, each of which can push goods in any direction, used to replace traditional conveyor lines for sorting. cellumation keeps its brand and operates independently, and its two founders stay on. This is Symbotic's next acquisition after ARMS Innovations and Fox Robotics.Source: Robotics & Automation News source

Yundie Technology (Chinese robotics startup) ｜ Series B+ tranche ｜ Several hundred million yuan · embodied

Investors include StellarS Capital, Asia Allied Infrastructure (00711.HK), Changzhe Investment, and others. This Guangzhou company entered through business-facing scenarios such as hotels, property management, medical and elderly care, and campus inspection; its products fall into two lines: the RoboForge embodied brain and the WING task-oriented robot body. The funds will go toward improving the reliability of sustained robot operation and cross-scenario deployment efficiency.Source: Robot Dajiangtang source

DareRobot (帝尔博格) ｜ Pre-A+ round ｜ Hundreds of millions of yuan / nearly 100 million yuan · industrial

Fosun RZ Capital, Qingshi Asset Management, Zhengxi Capital, Xigaotou, Bozhong Precision, and Suchuangtou participated. The two sources disagree on the amount: Robot Dajiangtang writes "hundreds of millions of yuan," while Cyzone writes "nearly 100 million yuan." The company makes cutting, grinding, welding, and coating robots for heavy-industry settings such as defense, steel, shipbuilding, and energy equipment. CEO Zhang Ben says its grinding robot has been running stably on the production line of a large steel company for over a year.Source: Robot Dajiangtang source; Cyzone source

NEXFORM (新元体) ｜ Seed round ｜ Tens of millions of yuan · embodied ⚠️ Single-source claim

Investors include Gobi Ventures, Zhejiang University Youchuang, and others. The team comes from the Robotics Institute of Zhejiang University; the product is general-purpose robot hardware plus a developer toolchain. The company says it has signed with multiple hotel groups and property management companies, with cumulative contracts worth hundreds of millions of yuan, an order of magnitude larger than the seed round itself.Source: Pedaily (投资界) source

An auto parts plant in Xiaoshan, Hangzhou, deploys nearly 20 humanoid robots for loading and unloading · humanoid

Over the past year or more, this plant has progressively put nearly 20 industrial humanoid robots into service, handling loading and unloading on assembly lines. They can now reliably handle more than a dozen specifications of auto parts, at 20 to 25 seconds per cycle. The plant's head makes a preliminary estimate that the labor cost saved will pay back the investment in two to three years; the next step is to have the robots also take photos during workshop inspections. The report does not disclose the name of the company or the robot manufacturer.Source: China Fortune (China News Service) source

Kexing Shikong (刻行时空) begins delivering a 200-million-yuan overseas real-robot data order · adjacent ⚠️ Vendor claim

The contract is worth over 200 million yuan, covers about 300,000 hours in total, and the customer is overseas; the first batch of more than 1,000 hours of multi-embodiment real-robot data has passed acceptance. The data was collected with Unitree robots and includes whole-body control and long-horizon continuous manipulation; data ownership was confirmed and compliant cross-border delivery was completed in the Lingang New Area. Jiemian News published this piece marked as reprinted "for advertising purposes"; the customer's name and unit price are both undisclosed.Source: Jiemian News source

Uber begins testing robotaxis in Manchester, with vehicles from Pony.ai and Verne · autonomy

The current testing maps streets with safety drivers and no passengers. Transport for Greater Manchester says it has not received any operating application. Pony.ai says its plan to deploy more than 2,000 robotaxis in five European cities is being "unlocked step by step." Uber expects to offer robotaxi service in up to 15 cities worldwide by the end of this year, mainly in the US and the Middle East. Driver unions GMB and ADCU both say this will put thousands of drivers out of work.Source: The Guardian source; Guandian (观点网) source

Waymo has imported over 3,200 Zeekr CM1-E vehicles, with per-vehicle cost above $100,000 · autonomy ⚠️ Compiled figures

This Autohome piece is compiled from customs bill-of-lading data: since 2024 Waymo has imported a cumulative over 3,200 Zeekr CM1-E vehicles, over 2,600 of them entering through the Port of Los Angeles this year; named Ojai in the US, more than 100 are already operating in Los Angeles and San Francisco. The cost is the article's own estimate: a declared customs value of about $38,000, plus 127.5% in tariffs for a landed cost of about $86,500, plus about $25,000 for the sixth-generation autonomous driving kit, for over $100,000 per vehicle. Even so, that is only about half the cost of the older Jaguar I-PACE conversions (over $200,000). To comply with US connected-vehicle rules, the imported vehicles come without autonomous driving and connectivity functions; after arriving at the Mesa, Arizona plant, Waymo's sensors and compute platform are added together with Magna.Source: Autohome source

UBTech's consumer humanoid U1 has begun small-batch delivery, with a cheaper model coming in Q4 · humanoid

UBTech (Chinese humanoid maker) Vice President Jiao Jichao says the U1 began small-batch delivery to end customers in September, factory capacity is ramping up, and a lower-priced model such as a "youth edition" will launch in Q4. When the U1 was unveiled in June it came in three tiers: the half-body Lite at 119,800 yuan, the full-body Pro at 169,800 yuan, and the Ultra at 880,000 or 990,000 yuan; the company then announced that all-channel orders had passed 13,000 units. How many units were delivered in this small batch has not been disclosed.Source: 21st Century Business Herald source

Google DeepMind releases Gemini 4 Argon, one model covering robot control, simulation building, and data labeling · world-model

Gemini 4 Argon is a general physical-reasoning model DeepMind released on October 14 (Beijing time); the company says it scored 87% on pure reasoning on the ERQA benchmark, which tests 3D physical and spatial reasoning, beating all other frontier models. ⚠️ Vendor claim The team showed it a demonstration video of a beanbag toss and had it guide an Apptronik Apollo humanoid through repeated practice, ending with 7 consecutive hits; the uses shown on the developer blog, such as controlling Spot, reconstructing simulation scenes, and breaking videos into tasks, are likewise officially chosen demos, which are capability demonstrations and not the same thing as production readiness. The same day, Aaron Saunders, DeepMind's Vice President of Embodied Systems and Robotics, announced a new robotics lab in Cambridge, Massachusetts.Source: Google Developers Blog source; Phoenix New Media (reposted from Robot Qianzhan) source; AI Insider source

China releases the first two national standardization guidance technical documents for embodied AI, both targeting training data · adjacent

China's State Administration for Market Regulation and Standardization Administration recently released GB/Z 220-2026, Artificial Intelligence — Technical Requirements for Embodied AI Data Generation Platforms, and GB/Z 218.1-2026, Artificial Intelligence — Embodied AI Data Quality Specification — Part 1: Real Data. The former gives a three-layer reference framework of engine, generation, and application layers, steering the industry toward a data supply route of "a small share of real-robot collection plus large-scale simulation synthesis." The latter assesses real datasets on both process quality and content quality. Both documents are administered by China's national AI standardization committee (全国信标委), with more than 50 organizations, including the China Electronics Standardization Institute and Huawei Cloud, participating in drafting, and the National and Local Joint Engineering Center (国地中心) took part in development and validation. GB/Z documents are guidance technical documents and are not mandatory.Source: Gasgoo source

China's customs authority: industrial robot exports in the first three quarters nearly 10 billion yuan, up 16.9% year on year · industrial

Wang Jun, Deputy Director of China's General Administration of Customs, announced at a State Council Information Office press conference on October 14 that China's industrial robot exports in the first three quarters were nearly 10 billion yuan, up 16.9% year on year. At the same briefing, ship exports were reported at over 350 billion yuan, so industrial robot exports are less than one-thirtieth of that.Source: 21st Century Business Herald source

The top five humanoid robot companies by first-half revenue are all Chinese; two research firms' unit counts differ by more than 10,000 · humanoid

The full-size humanoid share from this Frost & Sullivan report was covered the day before yesterday; its revenue ranking is as follows: in the first half of 2026, UBTech ranked first with 620 million yuan (16.5%), Zhiyuan Robotics (AgiBot, Chinese humanoid startup) 610 million yuan (16.1%), Unitree (Chinese robot maker) 580 million yuan (15.3%), Galbot 210 million yuan, and Leju Robotics (Chinese humanoid maker) 150 million yuan; the five together account for 57.4% of global revenue, and the figures exclude vendors' internal and related-party transactions. The report puts first-half global sales at 36,000 units and revenue at $540 million. The IDC figures cited in a Booster Robotics press release are: global first-half shipments of nearly 25,000 units and a market size of over $740 million, with Booster among the global top three by shipments. The two firms' unit counts differ by more than 10,000, while the dollar figure is higher at IDC.Source: 21st Century Business Herald source; FinancialContent source

NEURA releases the multimodal foundation model Neuron AI, adding touch, sound, and memory · world-model

Neuron AI, released by Germany's NEURA Robotics, uses a "perception-language-action" (PLA) architecture. It aligns 3D vision, touch and force, audio, robot state, and language into one temporal space before converting them to actions; a memory module lets the robot remember objects that have moved out of view. Founder and CEO David Reger says: "An intelligent robot needs a brain that brings all the senses together and remembers what has happened." The release gives two examples: confirming a lock is open by sound and touch, and judging weld quality by the sound of welding.Source: AFP (Business Wire wire copy) source

Nature Will (Beijing embodied AI company) releases Natura-1, driving dexterous hands with a million hours of human data · embodied ⚠️ Vendor claim

Beijing embodied AI company Nature Will released Natura-1. It is a 50B-parameter MoE model pretrained on 1 million hours of first-person human experience data; taking language and vision as input, it drives a 20-degree-of-freedom dexterous hand within a 100-millisecond control cycle. The company released an uncut, real-speed video 1 minute 12 seconds long in which both hands completely wrap a single wonton. Its headline selling point is executing new tasks it has not practiced, without extra adaptation; for now, there are only the company's own demos.Source: 36Kr source

Former Ford EV chief Doug Field becomes president of Rivian-affiliated Mind Robotics · industrial

Mind Robotics is an industrial AI robotics company Rivian incubated last year. Doug Field becomes president, responsible for product, engineering, and operations, with Rivian CEO RJ Scaringe remaining involved. Field has served as head of Apple's Project Titan and as Tesla's senior vice president of engineering. Mind Robotics has been backed by over $1 billion from investors including Accel and a16z; its robots do not aim to replicate the human form, and Rivian's factory is its first customer as well as its model-training ground.Source: TechCrunch source

Wang Xingxing: "I underestimated AI"; "when GPT-7 comes out next year, it may directly crack the embodied AI problem" · humanoid ⚠️ Speech remarks

Unitree founder Wang Xingxing said at the 26th China International Industry Fair: "The biggest mistake I've made over the past dozen or so years is that I didn't believe in AI enough." He revealed that the team deployed GPT-6 on several kinds of robots for testing, and "the results were actually very good." At the World Robot Conference in August, his assessment of industry breakthroughs was still "two to three years at the fastest, five to ten years at the slowest." Unitree is internally building a "physical AI self-evolution" system in which AI takes part in writing code, training, deployment, and testing.Source: Dianshangpai (电商派) source