Key Takeaways

· Unitree-rival Deep Robotics (Chinese quadruped robot maker) updates its IPO prospectus: only 4 humanoid robots sold in two years combined, and a swing from profit to loss in the first half of this year

· Yinghe Robotics (Chinese robotics startup), once valued at $500 million, enters bankruptcy proceedings; first creditors' meeting set for October 19

· Denmark's largest retailer is trialing a Unitree-built humanoid robot (Chinese robot maker) for restocking in a supermarket; at this stage every movement relies on human teleoperation

· When the environment changes mid-action, all 14 manipulation policies see success rates fall, by 11.0 to 25.7 percentage points

· Without having a robot arm execute anything, and just comparing against demonstrations from a handheld gripper, VLA post-training raises the average progress score to 94.8

LIBERO-MAX: Changing the environment mid-action to see whether a VLA can finish the job · benchmark

This paper was selected for an oral presentation at the NeurIPS '26 Robotics World Modeling Workshop, with authors from Stanford, CMU, MIT, UT Austin and others. Existing robustness evaluations mostly change conditions before a task starts; LIBERO-MAX changes them midway through execution. The same task is run twice, with identical state, instruction, random seed and actions up to the point of change. One run then continues as normal, while the other has one of 8 types of events inserted, such as a target moving, a camera view switching or sensor noise. This makes it possible to tell whether a failure was caused by the disturbance or whether the model simply could not do the task in the first place. Across 14 policies, 224,000 simulation runs were completed, covering VLAs, hybrid architectures and world action models. Success rates fell for all of them, by 11.0–25.7 percentage points. Even the best performers under dynamic conditions, MolmoAct2 and π0.5, reached only 66.9% and 65.7%. Shortening the interval between action executions and re-observing more frequently also failed to close the gap.

Yunbei Zhang and Zijian Jin, co-first authors (Stanford · CMU · MIT · UIUC · UT Austin and others; collaborators include Peter Stone and Marco Pavone) · NeurIPS '26 Robotics World Modeling Workshop Oral · Coverage: Sohu source

HIL-UMI: Targeted VLA post-training without repeated robot-arm execution · vla

Post-training methods of the HG-DAgger kind require a person to stand by the robot arm and step in to correct it once it makes a mistake. Dong Hao's team moved this step onto the handheld-gripper UMI: while a person demonstrates, the model predicts actions in parallel, and when the two diverge beyond a threshold, the system prompts the operator to record that segment again; training then uses advantage-conditioned behavior cloning. The experiments used π0.5 as the base model and tested four tasks on a Franka robot arm. With the full method over three rounds, the average task progress score rose from 62.3 to 94.8; with the same number of collected frames, conventional SFT reached only 58.0. On a tabletop-tidying task, collection efficiency was 5.63 times that of HG-DAgger.

Zimu Han, Yiming Zeng et al. (Dong Hao's team, jointly with Qiyuan Robotics, Peking University, Xi'an Jiaotong University and others) · arXiv 2609.20659 source · Coverage: Geekpark source

AWM: A world model that predicts first, then infers the cause · world-model

Baize Tongjing (Hefei-based AI startup), a startup in Hefei, proposes the abductive world model AWM. The model first predicts the next frame, then infers which objects, what motion and which contacts produced that change; training can use unlabeled video directly. The paper reports a 75% success rate on Push-T, with AWM using 2,000 trajectories versus 14,728 for Causal-JEPA; when training data is raised to the same 14,728 trajectories, AWM reaches about 92%. The code is open source, and all of the above figures come from the team's own paper.

Baize Tongjing (Hefei; founder Bai Yinqi) · Paper: "Abductive World Modeling via Causal Representation Learning" · Coverage: Sohu source

Pangbote ｜ Series B ｜ several hundred million yuan · adjacent

Led by Shanghai Guotou Pioneer Fund, with participation from Hejun Puhua, Pudong Venture Capital and others, and an additional investment from existing shareholder Lanchi Ventures; Gaogu Capital served as exclusive financial adviser. Pangbote (Chinese tennis training hardware maker) makes autonomous tennis-training hardware driven by a sports foundation model. The funds will go to R&D on an AI sports foundation model, motion vision and training decision-making, while also expanding overseas channels and launching new products for more sports categories. Source: The Beijing News source; 36Kr source

Moxian Technology (Dongguan) ｜ Series B ｜ amount undisclosed · hardware

The investors are the XBOTPARK fund and AI² Robotics. Moxian makes flexible thin-film pressure sensors, and its FLX series does not use a traditional FSR structure. One of the investors is the embodied-AI company X Square Robot, making this a downstream complete-robot maker investing in an upstream pressure-sensor supplier. Source: Sohu source

Denmark's Føtex supermarket trials a humanoid robot for night-time restocking, fully teleoperated for now · humanoid

At a Føtex supermarket near Copenhagen, a humanoid robot has been practicing shelf stocking since June. The software comes from Danish startup Hive Robots, and the robot body is made by Unitree. Every movement is currently guided by a human operator, with the aim of teaching the AI to handle different products. The goal is to restock independently at night after closing, around early 2027. Føtex says employees spend 40% to 50% of their time on restocking. The robot is currently slower than a person and needs someone watching closely; for it to work autonomously among customers, Denmark would first have to change its laws. The pilot is part of a larger trial by Salling Group, Denmark's largest retailer, which has more than 2,000 stores across 6 European countries. Source: Africanews source; AP News source

UBTech and C&D Inc. form a joint venture, targeting 10,000 humanoid robots sold by 2030 · humanoid ⚠️ Planned figure

UBTech (Chinese humanoid robot maker) and C&D Inc. (Xiamen-based state-linked trading conglomerate) signed an agreement in Xiamen on October 9 to establish a joint venture, with a target of selling no fewer than 10,000 UBTech humanoid robots combined by 2030. C&D said in a filing that it will push for robots to enter government exhibition halls, commercial complexes and industrial parks. At the end of September, UBTech's CFO said the U1 could actually deliver only 1,500 to 2,000 units this year (previously reported). Source: 36Kr source; Eastmoney source; Jiemian News

Tesla officially says Austin has over 300 Cybercabs; Dallas is next · autonomy

Texas registrations had previously risen to 319 vehicles, and now Tesla's official Robotaxi account has also confirmed 300-plus Cybercabs operating in Austin, an eightfold increase in 30 days. Tesla AI vice president Ashok Elluswamy wrote on X "Dallas soon," without giving a date. NHTSA has put 21 questions to Tesla about how the Cybercab is certified and about its expansion plans, with a response deadline of October 30. Source: Not a Tesla App source

Deep Robotics updates its prospectus: 4 humanoids sold in two years, loss in the first half · humanoid

The DR-series humanoid robot brought in revenue of RMB 1.1711 million in 2024 and RMB 823,000 in 2025; only 4 units were sold over the two years combined, and humanoids made up 0.24% of 2025 revenue. Revenue comes mainly from quadrupeds; the average unit price of the flagship Jueying X series fell from RMB 287,500 to RMB 228,200, and gross margin fell from 54.35% to 39.75%. In 2025 Deep Robotics made its first profit, RMB 28.684 million, of which only RMB 15.1232 million was profit excluding non-recurring items. Revenue in the first half of 2026 was RMB 259 million, up 117.67% year on year, with a net loss attributable to shareholders of RMB 8.8063 million, a loss of RMB 18.5983 million excluding non-recurring items, and operating cash flow turning negative. The company's stated reasons include R&D expenses rising to about RMB 126 million and large-volume deliveries of low-margin models. The prospectus adds a new risk warning: "Customer A1" has already moved into similar products through in-house development and acquisitions. Humanoid production capacity is expected to reach only the hundreds of units in 2026; this STAR Market IPO aims to raise RMB 2.503 billion. Source: Ifeng (Dayu Finance) source

Yinghe Robotics enters bankruptcy proceedings; first creditors' meeting on October 19 · adjacent

About two months ago, the Shenzhen Intermediate People's Court issued a notice accepting Yinghe Robotics' bankruptcy and calling for creditor claims, with the first creditors' meeting set for October 19. Yinghe was registered in Shenzhen in 2020 and is 55% owned by Yingfeng Group (Chinese conglomerate); Yingfeng's controlling shareholder He Jianfeng is the son of Midea founder He Xiangjian. The company brought in Panda Capital, Gobi Partners, Meituan's strategic investment arm and ByteDance's strategic investment arm in turn, reaching a valuation of $500 million in 2021. It ran five business lines at once, including emergency management, urban governance and warehouse logistics, and industry media reported that most orders stayed at the demonstration stage; there has also been no news of production at its planned RMB 10 billion base in Shunde. Qichacha records 58 judicial cases, 33 of which were closed as terminated enforcement. Li Lun, founding partner of Panda Capital, publicly alleged that executives drew high salaries and that Yingfeng cashed out before the trouble surfaced; these claims are unconfirmed, and neither Yingfeng nor He Jianfeng has responded. Also in September, pool-cleaning robot company Zhicheng Power entered bankruptcy review, while Shenlan Robotics (Chinese AI robotics company) had already been accepted for bankruptcy liquidation in May. Source: Sina Finance (Rongzhong Finance) source

San Francisco's fleet week puts Robotaxi no-go zones in place · autonomy

During the daytime on October 9 to 11, San Francisco's transportation agency used geofencing to keep Robotaxis out of the northern waterfront for passenger pickups and drop-offs, and the entire Presidio was shut down to them; Uber, Lyft and taxis were not restricted. Waymo said the avoidance zones were set up by its own initiative in coordination with the city and police. Multiple Robotaxis stopped on the road after the July 4 Golden Gate Bridge fireworks show, and there were also problems during the citywide power outage in December 2025; when Mayor Daniel Lurie asked the state to tighten regulation, he cited exactly those two incidents. Source: CBS News source