Today's Highlights

· XPENG (Chinese EV maker) names its robotaxi "YOYO" and opens beta testing, with a Guangzhou pilot fleet of nearly 100 vehicles

· Tesla renames FSD in Europe to TAD (Assisted Driving); the EU-wide vote is pushed to December at the earliest

· RoboSense (Chinese lidar maker) sold 145,200 lidar units into robotics and other fields in Q3, up 309% year on year

· ARC raises π0.5's real-robot task success rate by 82.2 percentage points without adding a single robot demonstration

· 93.2% of VioLA's training frames come from humans, with 100% real-robot zero-shot success on locomotion commands

· Guangdong sets a target of 80,000 humanoid robots produced per year by 2028

ARC: a reasoning recipe for robot foundation models · vla

Robot foundation models usually get stronger through bigger models and more demonstrations. ARC changes only how inference is done and collects no new data. The reasoning chain must land on the next action, spelling out why this step is appropriate and what effect it is expected to produce; these annotations are generated automatically from existing DROID demonstrations and then used to fine-tune π0.5 and Cosmos3-Nano-Policy. On real robots, π0.5's task success rate rises by 82.2 percentage points, and on RoboLab-Reasoning-50 the gain reaches up to 50 percentage points.

Gokul Puthumanaillam et al. · arXiv 2610.12386 source

VioLA: training a generalist humanoid policy on human data, walking zero-shot on real robots · locomotion

Generalist humanoid policies have so far needed task-by-task fine-tuning on teleoperated demonstrations before running on real robots. VioLA changes the prediction target: the policy outputs latent motion variables for the body and hands, which a pretrained controller executes, and human and robot motions are encoded into the same latent space, so human video can serve directly as training labels. Of the 140.6 million frames in the training pool, 93.2% come from humans; on locomotion commands the real-robot zero-shot success rate is 100%, versus 16.7% for GR00T N1.7 and 0% for Ψ0. Manipulation tasks reach 88.6% without targeted fine-tuning, and the code and weights are said to be forthcoming.

Mert Albaba et al. · arXiv 2610.12435 source

What 30,000 hours of first-person video cannot teach a world model · world-model

The authors scale first-person human video up to 30,000 hours (over 1,000 scene types, 14,000 collectors) and directly measure fidelity of human-object interaction in out-of-distribution scenes. With 100 times more data, both improve, but human-body modeling is close to saturation while object fidelity is much lower and improves more slowly. A supervision scheme that shifts model capacity from scene appearance toward object dynamics narrows the gap, though a large shortfall remains.

Jiahua Dong et al. · arXiv 2610.12464 source

DreamTrue: counterfactual post-training so robot world models follow actions accurately · world-model

Robot datasets contain few failed interactions, so world models are biased toward predicting success. DreamTrue rewrites recorded action trajectories to generate future videos with more contact scenarios, then post-trains with reinforcement learning using a video reward model trained on human-annotated defects. On AgiBot data, the human-evaluated interaction defect rate falls from 48.12% to 6.25%, taking first place in the world-model track of the AgiBot World Challenge 2026; HF 33↑.

Junyan Li et al. · arXiv 2610.12468 source

SuperNav: an agentic navigation system with no navigation fine-tuning · autonomy

HF 62↑, the highest community buzz today. SuperNav does not fine-tune a multimodal large model for navigation; it only adds an agentic wrapper, including navigation skills, physical-interaction tools, and task-progress management. The model points to the destination directly on the image and corrects itself based on execution feedback. It beats 4 baselines on instance-level, multi-goal, and demand-driven tasks, and has been deployed on a real quadruped robot.

Jinkai Zhang et al. (Zhejiang University) · arXiv 2610.12126 source

Embodied Turing Machines: no model at test time, policy made of pure code · manipulation

COAP uses code to track the robot, environment, and task state from images and proprioception, and all decisions are made by code; the shared codebase is repeatedly rewritten and expanded by a coding agent. On RoboDojo's 42 bimanual tasks, this codebase achieves a success rate of 70.24%, with no VLM or VLA in the test-time loop. When Robotera (Chinese humanoid startup)'s VPP2 topped the same benchmark at 32.26%, it stated that it did not use Agent RSI; this paper happens to take the RSI route, so the two scores should not be compared directly. HF 6↑.

Kairui Hu et al. · arXiv 2610.12369 source

SpatialHarness: adding rendered viewpoints so a frozen GPT-6 Astra can do fine manipulation · manipulation

Frontier multimodal models such as GPT-6 Astra can already control robots directly, but often fail at fine manipulation. The authors argue the problem is that cameras fail to capture key spatial relationships clearly, so they maintain online a simulated scene synchronized with the real robot and render supplementary viewpoints for the frozen model to see, without changing the policy or adding cameras. With the same GPT-6 Astra, plug-insertion success rises from 26.7% to 66.7%, and Tower of Hanoi from 0 to 100%.

Jiayu Wang et al. · arXiv 2610.12457 source

Reinforcement learning across a million parallel environments: sample at the capability frontier · manipulation

In large-scale parallel simulation, sampling task configurations uniformly means more and more compute is spent on configurations that are already mastered or still out of reach. SGS (success-guided sampling) concentrates training near the edge of the policy's capability, and with up to 2^20 (over a million) parallel environments it solves multi-terrain quadruped locomotion and contact-rich assembly that earlier methods could not; after distillation into an RGB policy, it transfers zero-shot to real-robot assembly.

Octi Zhang et al. · arXiv 2610.12465 source

Ultra (US) ｜ Series A ｜ $50 million ｜ valuation undisclosed · embodied

Led by Framework Ventures, with Y Combinator participating; together with the earlier $12 million seed round, the total is $62 million. Founded in 2024, Ultra makes the OP1, a fixed dual-arm robot rented by the month to e-commerce warehouses for packing, sorting, and kitting, and it is expanding its partnership with Physical Intelligence, which supplies the VLA model. The company says its robots have packed over 500,000 orders and handle up to 30% of fulfillment volume at Highline Commerce in Brooklyn. Sources: Crypto Briefing source; Dealroom source

Bensu Intelligence ｜ Seed ｜ amount undisclosed ｜ valuation undisclosed · world-model

Led by 5Y Capital (Chinese VC firm), with Ivy Capital, Linear Capital, and Huashan Capital participating. The company was founded in July 2026 by eight researchers from Tsinghua and Peking University backgrounds and builds action-native embodied foundation models. Its cross-embodiment model X-VLA won the IROS 2025 AgiBot (Chinese humanoid maker) Challenge among more than 400 teams, using under 1,000 hours of pretraining data. Source: Electronic Engineering Album (Hangjia Shuo Robots) source

Diffusion Intelligence (Shenzhen) ｜ two consecutive rounds ｜ several hundred million yuan combined ｜ valuation undisclosed · adjacent

Co-led by Matrix Partners China, Shunwei Capital, and Legend Capital, with Huawei Hubble, CAS Star, and Horizon Robotics (Chinese chip company) following on. Founded in May 2026, the company's founding team is Kong Lingpeng, co-director of the HKU NLP lab, and his PhD students Gong Shansan and Ye Jiacheng. The company builds diffusion language models (dLLMs, language models that generate in parallel and can go back and revise what they have already generated) and is training a 30-billion-parameter model, aimed at on-device agents and robot task planning. In adaptation tests with Acrab, the two sides say on-device agents ran 5 times faster. Source: Nandu Bay Finance Society source

Jianzhi Robotics ｜ Baidu and Galbot take stakes ｜ amount undisclosed · embodied

Tianyancha shows that the affiliated Jianzhi Xinchuang (Beijing) Robot Technology added Beijing Baidu Netcom and Suzhou Gaibote, a Galbot (Chinese embodied-AI startup) subsidiary, as shareholders, with registered capital rising from about 193,000 yuan to about 10.524 million yuan. Founded in May 2025, the company's business scope includes manufacturing and selling service and consumer robots. Source: Radar Finance source

Biren Technology ｜ strategic investment ｜ $515 million · hardware

The investors are UBS Asset Management and CICC (Chinese investment bank), disclosed by Yiou Data on October 8. Biren is a Shanghai-based high-performance computing solutions provider; total funding disclosed that day in the intelligent technology sector was about 4.077 billion yuan, of which this deal made up most. Source: Yiou Data source

XPENG robotaxi named "YOYO", nearly 100 vehicles in beta in Guangzhou · autonomy

On October 8, XPENG named its robotaxi business "XPENG YOYO" and launched a mini-program at the same time, gradually releasing beta slots to users. The vehicle is the robotaxi version of the XPENG GX, with 4 in-house Turing chips, 3,000 TOPS of effective compute, running a second-generation VLA. It obtained a Guangzhou road-testing permit in January this year, the first production vehicle rolled off the line in May, and beta testing began in July; as of August, Guangzhou had accumulated over 2,000 beta orders. XPENG vice chairman and president Gu Hongdi told CNA that the Guangzhou pilot fleet is nearly 100 vehicles, and "starting next year, we hope to launch pilots in some overseas cities." He Xiaopeng's goal is driverless passenger service with no safety operator in 2027, and the Guangzhou business could reach per-vehicle breakeven as early as the second half of 2027. As a reference point, the same report notes that Apollo Go (Baidu's robotaxi service) has passed 22 million cumulative orders globally. Sources: Wangtongshe source; CNA source

EACON Mining: over 600 driverless mining trucks enter coal mining · autonomy

At the end of September, EACON Mining (Chinese autonomous mining-truck firm) reported more than 1,500 electric driverless mining trucks, and it has now disclosed figures for coal mining scenarios. As of September, over 600 driverless mining trucks were mining coal at more than 10 open-pit coal mines in Xinjiang and Inner Mongolia, with cumulative coal haulage exceeding 200 million tonnes. At one Xinjiang mine, 24 trucks were first deployed in the winter of 2023 and the fleet has since grown to 111, serving 6 crushing stations, each handling 35+ truck dumps per hour, with efficiency on par with human driving. At CHN Energy's Zhaniher open-pit mine, 5 trucks of the 90-tonne class saw average daily load per truck rise from 4,500 tonnes to 7,200 tonnes, with human interventions falling to 1.3 per 100 kilometers. Source: China Nonferrous Metals News source

Exotec Skypod runs for a year at First Supply, next-day orders up to 20% of daily volume · industrial

US HVAC and building-materials wholesaler First Supply has used an Exotec Skypod system at its West Salem, Wisconsin distribution center for over a year, and orders placed before 5 p.m. can be delivered by the next morning. Such next-day orders make up as much as 20% of the center's daily order volume. Chief Logistics Officer Steve Baer said Skypod lets the company handle peak season and daily order peaks. Source: Robotics & Automation News source

Neolix's Shenzhen night deliveries expand to 437 routes · autonomy ⚠️ vendor-reported

After Shenzhen opened night-time road access to autonomous delivery vehicles in March, Neolix (Chinese autonomous delivery-vehicle maker)'s night deliveries have expanded to 437 routes, and Bloomberg ran a feature using it as an example. Founder Yu Enyuan told Bloomberg: "Autonomous delivery vehicles do the work people don't want to do, are unwilling to do, or cannot do." The company says it operates over 27,000 vehicles in 300 cities across 15 countries. Sources: Jinrongjie source; Bloomberg source

Pony.ai and Uber bring seventh-generation robotaxis to London · autonomy ⚠️ planned figures

Following Zagreb, the two announced on October 8 that they are advancing a London project, with road testing expected to begin in the coming weeks. Under the two companies' August plan, over 2,000 robotaxis are to be deployed across 5 European cities. Source: Gasgoo source

Skild AI quadruped robots on a six-month trial at Pittsburgh airport · world-model

Through its xBridge innovation program, Pittsburgh International Airport has 4 quadruped robots running Skild AI software patrol the terminal. They first test air quality, and later look for floor spills and overflowing trash bins. The pilot lasts six months, with Skild staff stationed on site to monitor. Sources: International Airport Review source; Technical.ly

Tesla renames FSD to TAD in Europe, EU vote postponed · autonomy

Tesla's German website has already renamed FSD (Supervised) to TAD (Tesla Assisted Driving), with the Belgian and Danish sites adjusted in step. Germany's Federal Ministry of Transport stated on October 6 that the renaming was proposed by Tesla, in response to possible misunderstandings caused by the FSD name; Germany supports moving the system into the EU approval process. The EU-wide vote originally scheduled for October 6 has been postponed to December at the earliest, and Tesla needs the support of at least 15 member states and a specific majority. The Netherlands, Lithuania, Estonia, Belgium, Denmark, Slovenia, the Czech Republic, and Croatia, 8 countries in all, have approved the relevant functions. In the China market the name "Tesla Assisted Driving" was already adopted in May, and the rollout timing is still undetermined. Source: National Business Daily source

Guangdong sets a target of 80,000 humanoids a year by 2028 · humanoid

Wu Hong, first-level inspector at the Guangdong Provincial Department of Industry and Information Technology, said at the province's advanced manufacturing conference on the 9th that Guangdong aims to push annual humanoid robot output past 80,000 units by 2028, with AI adoption among above-scale manufacturers exceeding 50%. She expects Guangdong to account for about one-third of China's humanoid output in 2026, and the number of humanoids "working" in factories this year could exceed 3,000. IFR figures put global sales of full-size commercial and professional humanoids in 2025 at about 7,000 units; that is sales while this is production, so the bases differ, and 80,000 units is about 11 times that figure. UBTech (Chinese humanoid maker) chairman Zhou Jian said at the conference: "The value of humanoid robots lies not in looking like humans, but in being able to enter factories, pick up tools, and complete real work." Sources: China News Service source; Sina Finance source

UBTech signs strategic cooperation with FAW-Volkswagen for in-plant logistics · humanoid

On October 8, UBTech announced a partnership with FAW-Volkswagen (Chinese joint venture) to develop and test application scenarios for embodied-intelligence robots in logistics; the specific form of cooperation was not made public. The two had already cooperated in 2024 at FAW-Volkswagen's Qingdao branch, when the Walker S Lite entered the factory for training on vehicle quality inspection. UBTech's 2025 full-size humanoid revenue was 820 million yuan, with sales of 1,079 units. Sources: Sina Finance source; Gasgoo source

BYD's humanoid robot "Xiaodi" gets design patent grant · humanoid

Design patent CN 310247768 S was published by the China National Intellectual Property Administration, with BYD as patentee, a grant announcement date of October 6, and the product name "embodied intelligence robot". This is the first formal public release of the appearance of BYD's in-house humanoid, corresponding to the model codenamed "Xiaodi": 1.61 m tall, 58.5 kg, with 31 degrees of freedom across the body, positioned for commercial in-store service. BYD executive vice president Li Ke said earlier that the goal is to place 2 to 3 units in each store. Source: Kuaikeji source

Humanoid daily rental falls below 1,000 yuan while industrial robots raise prices · humanoid

According to the STAR Market Daily, humanoid robot daily rental fees have fallen from the tens-of-thousands-of-yuan range in early last year to 2,000-3,000 yuan at the start of this year, and have now dropped below 1,000 yuan; on Xianyu, some merchants list Unitree humanoid rentals starting at 289 yuan/day. Complete-unit selling prices are also trending down: Noetix Robotics' "Bumi" has an official price of 9,998 yuan, and the Unitree R1 has fallen from 39,900 yuan to 29,900 yuan. Unitree's prospectus shows its humanoid unit price fell from 593,400 yuan in 2023 to 166,400 yuan in 2025. Industrial robots are moving the opposite way: Inovance (Chinese automation maker) has raised prices on SCARA and articulated robots by 5% to 15% since March, and Efort (Chinese industrial robot maker) has raised prices twice this year, citing higher prices for copper, aluminum, and memory chips. Sources: Jinrongjie (STAR Market Daily) source; Phoenix WEEKLY Finance source

Over 70 training grounds; report estimates selling data takes 3+ years to pay back · adjacent

The "Embodied Intelligence Training Ground Research Report (2026)", released by CAICT (China Academy of Information and Communications Technology) and other bodies, counts over 70 embodied training grounds in operation across China as of the end of June, with another 46 in planning or construction, 86% of them covering industrial manufacturing scenarios. Building a comprehensive training ground of several thousand square meters can cost tens of millions to over 100 million yuan. The report estimates that with 50 million yuan invested and 100,000 hours of data produced and fully sold each year, selling data alone would still take over three years to pay back. Overseas robot-data company Eidon AI has ceased operations. Founder Sam Padilla reflected that the team took about five months to obtain about 3,000 hours of data, and that "what labs need is not a one-off dataset, but a reliable, continuous data supply." Source: The Paper (21st Century Business Herald) source

UK allocates £40 million for 8 robotics adoption hubs · adjacent

The UK government announced 8 robotics adoption hubs across the country, with £40 million in total funding, to help businesses and public bodies adopt automation. Among them, the Newcastle hub focuses on robot-assisted surgery and the Lincoln hub on agriculture and food. A further £2.5 million supports 12 robotics training and apprenticeship projects. AI and Robotics Minister Kanishka Narayan said that using robots across industries can "accelerate the UK's reindustrialization". Source: BBC source

Helm.ai says it signed $70 million in contracts in 12 months · autonomy ⚠️ vendor-reported

Autonomous-driving foundation-model company Helm.ai announced that over the past 12 months it has signed $70 million in commercial contracts with global automakers, Tier 1 suppliers, and industrial automation companies, and says the company is moving toward operating breakeven. Contract details and the customer list were not disclosed. Source: Business Wire (FinancialContent) source