Key Takeaways

· Electric driverless mining trucks from EACON Mining (Chinese autonomous mining-truck maker) top 1,500 units, up from 800 in March

· UBTech's U1 reports 13,361 units in orders, but its CFO says only 1,500 to 2,000 units can actually be delivered this year

· According to CNBC citing three people familiar with the matter, China's securities regulator has set three thresholds for humanoid robot IPOs

· An audit found 22 bugs across 7 VLA simulation benchmarks; once fixed, the ranking of methods flips

· The dual-arm VLA Rho is open-weight, and 15 corrective demonstrations are enough for online adaptation

· Singapore has opened a humanoid robot public safety centre, aiming to have humanoid robots working alongside police officers on real missions by 2028

VLA acceleration that seems to "get more accurate the faster it runs" is most likely a benchmark problem · benchmark

Some training-free VLA acceleration methods only approximate the original model's computation, yet measure a higher success rate than the original model. The authors traced this anomaly and found 22 bugs and 4 design flaws across 7 benchmarks, including RoboTwin, LIBERO-Plus and VLABench. After the bugs were fixed, the original model on one task rose from last place to first; after the flaws were corrected, the original model on another task went from trailing the accelerated method by 21 percentage points to leading by 5 points. The fix code and the modified benchmark settings have been released together.

Qiwei Chen et al. · arXiv 2609.37771 source

Rho: an open-weight VLA for three dual-arm embodiments · vla

Rho targets three dual-arm platforms common in labs and industry, YAM Box, UR AI Trainer and FR3 Duo, and releases a base model plus a dedicated checkpoint for each embodiment. Experiments show that adding a stage of "embodiment mid-training" improves downstream adaptation. The model comes with an online adaptation module: it freezes the flow-matching action expert and trains only a small policy that selects the noise input, and 15 corrective demonstrations are enough to handle cases at the edge of the offline fine-tuning distribution. The authors say it matches or exceeds existing open-weight VLAs on the tasks tested.

Rho Team · arXiv 2609.38164 source

EVO-WAM: a world action model that self-evolves on its own generated trajectories · world-model

This work collects no new demonstrations and does no trial and error in the environment. The world action model generates video and actions itself; a VLM first picks out the clips that complete the task, then an inverse dynamics model checks whether the video and actions are consistent, and finally the clips that pass the check are used for iterative training. On 7 unseen RoboTwin 2.0 tasks, Cosmos3's average success rate rises from 26.9% to 68.0%. On real robots, 3 long-horizon compositional tasks rise from 20.0% to 76.7%. HF 15↑.

Shiyang Zhou et al. · arXiv 2609.38057 source

PRISM: expanding a few real human videos into hundreds lets a humanoid learn to carry objects with zero real-robot fine-tuning · locomotion

Human-object videos with the full body clearly visible and unoccluded interaction are hard to collect at scale, and this is the bottleneck for teaching humanoid loco-manipulation from video. PRISM uses video-to-video generation to expand a small number of real clips into hundreds of "counterfactual" interaction videos, then reconstructs the human and object motion and retargets it into physically feasible trajectories. The single policy trained this way is deployed without real-robot fine-tuning and, using only onboard depth, can pick up, carry and put down new boxes, buckets and balls. Pieter Abbeel and Jitendra Malik are among the authors.

Zihan Wang et al. · arXiv 2609.38172 source

LIBERO-MAX: when the scene changes mid-task, all 14 policies lose performance · benchmark

Most robustness benchmarks change external conditions only at the start of a run. LIBERO-MAX uses 8,000 paired test cases to move the target, switch the viewpoint or add obstacles midway through a task. As a result, the success rates of 14 VLAs, hybrid policies and world action policies fall by 11.0 to 25.7 percentage points, with geometric and observation changes being a shared weakness.

Yunbei Zhang et al. · arXiv 2609.36518 source

FP2: bolting a force-control layer onto robot foundation models · manipulation

FP2 splits the division of labour: the foundation model only generates task actions, while a separate high-frequency policy only regulates contact force, reading the foundation model's context to predict force-control parameters. Across 4 foundation-model backbones and 4 contact-rich real-robot tasks, both task performance and force-regulation quality beat the original models.

Hongjie Fang et al. · arXiv 2609.37433 source

V-JEPA Policy: building a world action model without inheriting a whole video-generation model · world-model

Mainstream world action models inherit a complete pretrained video generator. This work uses only the latent space of a frozen V-JEPA 2.1 encoder and trains a future-latent predictor and a flow-matching action expert on top of it simultaneously, with 0.9B total parameters, matching representative baselines on LIBERO and RoboCasa-GR1. Pretraining the predictor first on DROID video without action labels noticeably improves out-of-distribution generalization. The code is open-sourced.

Yang Zhang et al. · arXiv 2609.37250 source

CrossBFM: multiple humanoids sharing one behavior latent space · locomotion

A behavior foundation model takes hundreds of GPU hours to train a latent space for one robot, and switching to another robot yields a space that does not align with the original. CrossBFM uses a unified encoder with no robot-specific parameters to distill the behavior space onto multiple humanoids in under 1 GPU hour. On robots not included in training, it recovers up to 89% of tracking performance, and it was also validated on real robots. HF 12↑.

Tan-Dzung Do et al. · arXiv 2609.38087 source

A survey of in-context learning for robots · manipulation

Today's most popular paper in the community (HF 181↑). The survey sorts robot ICL into four categories by "how in-context evidence connects to execution": context-conditioned policies, geometric demonstration transfer, world-model-based control, and skill and agent execution. It also discusses how to evaluate in a way that separates "understood the demonstration" from "transferred to the physical world".

Haojian Huang et al. · arXiv 2609.36012 source

Sudo Technology ｜ New round ｜ over RMB 1 billion ｜ previously valued above US$2 billion · embodied

Shanghai Guotou Xianfeng (Shanghai state-backed investment fund), Futeng Capital and Guotai Haitong co-led, with existing shareholders including Xiaomi following on; Ningde, Alibaba, Tencent, Ant, and IDG had already invested earlier. Founded in 2025, the company develops the full stack in-house, from models, data and simulation to the robot body, and co-founder and CTO Su Hao is one of the leaders behind ShapeNet, SAPIEN and ManiSkill. The company says its Sudo R1 system uses no task-specific real-robot data and achieves near 100% zero-shot success on key tasks. This figure has not been independently reproduced. ⚠️ Vendor-reportedSource: Phoenix New Media (Robot Qianzhan) source

General Intuition ｜ New round ｜ US$220 million ｜ valued at US$6.2 billion · world-model

Valor Equity Partners, Khosla, General Catalyst and others participated. The company spun out a year ago from Medal, a game-clip recording tool, and trains world models on gameplay footage to generate synthetic training video. MIRA, released in June, has 5.6 billion parameters and can render 720×576 frames at 20 frames per second on a single B200. For now MIRA can only generate footage of one game and is a research demo; the commercial version is still being tested with a small number of robotics and simulation customers. Source: SiliconANGLE source

Efficient Computer ｜ Series B ｜ over US$97 million ｜ valued at US$650 million · hardware

Led by TQ Ventures, with Toyota Ventures and others following on, bringing total funding to US$173 million. The company says its Electron E1 low-power general-purpose processor is in mass production, and the new funds will go toward ramping up to its first customers and extending the architecture to data-center scale. Source: PR Newswire source

Inbolt ｜ New round ｜ €11 million · industrial

A Paris-based company, led by Shift4Good, with total funding of €30 million. It uses 3D vision plus a robot-brand-agnostic AI software layer to let industrial robots correct their motions in real time, and is installed on more than 200 robots at over 100 factories across 3 continents, with customers including Bosch, Ford, Stellantis and Toyota. Next it is moving into data centers and electronics manufacturing. Source: Tech.eu source; Dealroom source

Tesla ｜ Credit facilities ｜ US$30 billion · autonomy

According to an 8-K dated September 29, Tesla signed three new unsecured credit facilities: a US$20 billion three-year delayed-draw term loan led by Citi, plus an US$8 billion five-year and a US$2 billion 364-day revolving facility led by Wells Fargo, while terminating its 2023 US$5 billion revolving agreement. As of the signing date, none of the facilities had been drawn. Source: infohub.kz source; Tesla 8-K

Mengshi Technology ｜ Series A+ ｜ over RMB 100 million ｜ closed · autonomy

Liding Investment, Hangzhou Heda and Huifu Investment participated, and a Series B is underway. The company builds L4 systems for commercial vehicles in confined settings such as mining sites, and founder Qiu Changwu holds a PhD from Shanghai Jiao Tong University's robotics institute. Source: Sohu (Rui Finance) source

EACON Mining's electric driverless mining trucks pass 1,500 units, nearly doubling in half a year · autonomy

The count was 800 in March 2026 and now exceeds 1,500 units. As of early September, electric vehicles make up about 42% of its driverless fleet, diesel hybrids 41% and methanol hybrids 16%. China Energy Investment Corporation's Zhundong open-pit coal mine has 120 electric driverless mining trucks sharing 28 charging bays, with the dispatch system sending vehicles to charge on their own based on battery level, energy-consumption forecasts and the production plan. Zijin Mining's Julong copper mine has 60 units running routinely at high altitude, with attendance above 90%. After trial-running 6 units for more than a year, Baowu's Taihe iron mine added another 12. Source: Global Mining Review source

UBTech U1: 13,361 orders, and fewer than 20% deliverable this year · humanoid

At the June 30 launch event, CEO Zhou Jian announced 13,361 units in orders across all channels, with first deliveries set to begin on September 16. On the August 30 earnings call, CFO Zhang Ju said the U1 series could actually deliver 1,500 to 2,000 units this year. On September 28 the company stated that it was scheduling production in batches in order of reservation, yet customer service at its Tmall and JD.com flagship stores said the first batch was pushed to after November 30. Pre-sales collected only a refundable RMB 3,000 deposit, and how many orders will proceed to final payment will only become clear after the delivery window. In the first half, 921 full-size humanoid robots were sold, of which the Walker S2 accounted for more than 500 delivered in industrial settings. Source: Sohu (Xinghe Business Observer) source

Critical Point's OmniHand dexterous hands exceed 15,000 units delivered cumulatively · embodied

The company says that as of September, cumulative deliveries of the OmniHand series (excluding grippers) exceeded 15,000 units, used in industrial, logistics and service settings. In a Suzhou embroidery video released at the same time, the fully direct-drive OmniHand 3 Ultra completed basic steps such as twisting thread, splitting silk and threading a needle. The video is a capability demonstration and does not represent a mass-production scenario. ⚠️ Vendor-reportedSource: Sohu (Shanghai Securities News) source; PA Media source

AgiBot's first-half humanoid shipments exceed 8,600 units · humanoid

IDC's global first-half shipment data was reported yesterday (nearly 25,000 units). AgiBot (Zhiyuan Robotics, Chinese humanoid startup) self-reported first-half shipments on September 30 of over 8,600 units, claiming a 35% global share and more than 45% in China. ⚠️ Vendor-reportedSource: Guandian Wang source

Iron Mountain is first to deploy Dexory's new-generation 18-meter inventory robot · industrial

The new AMR scans 16,100 locations per hour and reaches shelves 18 meters high, where the previous generation only reached 12 meters. Iron Mountain says the system raised inventory accuracy from 97.8% to 99.9%. The two companies have worked together at 4 UK warehouses since 2024. Source: Robotics & Automation News source

Kodiak names IKEA as its first customer for driverless line-haul freight · autonomy

The driverless segment is a 219-mile route on Texas's I-45, connecting Kodiak's two terminals in Houston and Dallas, and is planned to begin running later this year. Over the past four years, Kodiak has hauled more than 1,300 loads for IKEA with safety drivers on board, totaling 750,000 miles. ⚠️ PlannedSource: bizjournals source; FreightWaves

ICE buys 4 modified Spot robot dogs for US$1.3 million · adjacent

The robot dogs were modified by Radeco of Connecticut from Boston Dynamics' Spot, with added robotic arms, thermal imaging and explosives sensors, at roughly US$110,000 each excluding the add-ons. ICE did not put the purchase out to bid. The contract states they may be used for immigration enforcement; Boston Dynamics spokesperson Sarah Tabak said "any claim that these robots will be used to assist with detentions or deportations is false," and said the contract wording is boilerplate. Source: The Boston Globe source

China's securities regulator reportedly sets three thresholds for humanoid robot IPOs · humanoid

CNBC, citing three people familiar with the matter, says the China Securities Regulatory Commission has privately required humanoid robot companies to meet up to three conditions before listing: sustainable revenue and commercial orders, a path to narrowing losses, and command of core technology such as the "brain" or dexterous hands. Some sources say meeting two of them may also be enough to pass. At least twenty-plus embodied AI companies have filed to list in Hong Kong. Unitree (Chinese robot maker) closed at RMB 845 on its first trading day and fell to RMB 459.65 this Monday. ⚠️ RumorSource: Tekedia source; CNBC source

DEEP Robotics' updated filing: its largest customer starts developing similar products in-house · embodied

After updating its STAR Market prospectus earlier this week, the updated filing discloses for the first time that a major customer, A1, is developing similar products through in-house R&D and acquisitions, with contracts worth RMB 55.075 million currently being performed between the two. Net profit attributable to shareholders in the first half was -RMB 8.8063 million, and the gross margin of the Jueying X series fell from 54.35% to 39.75%. The DR series humanoid robots sold only 6 units in the first half. Source: Securities Star source

Singapore opens a humanoid robot public safety centre · humanoid

At the opening of H2RC, Second Minister for Home Affairs Tan Chen Hui (typed 唐振辉 in the source) said AI humanoid robots will work alongside hazmat, prison and criminal-investigation officers on real missions by 2028. The centre received about 100 million (the original reads "$100 million") in investment, covers 29,000 square feet, and replicates a post-blast scene and prison cells on site. Models are trained on HTX's own computing power, and HTX is the first Singapore institution to put Nvidia's GB300 into use. Source: Yahoo News Singapore source

Fanuc and Yaskawa step up factory building in the US · industrial

Fanuc is investing US$90 million in a plant on the former General Motors site in Pontiac, Michigan, and Yaskawa is investing about US$180 million in Wisconsin. In the fiscal year ending this March, the Americas accounted for 27.1% of Fanuc's revenue, exceeding China's 26.6%. According to estimates from Japan's Okasan Securities, Fanuc's share of the US industrial robot market in 2024 was 56%; in China, it lost its top spot in shipments last year. Source: Nikkei Asia source

Yinghe Robotics, backed by Meituan and ByteDance, goes bankrupt · adjacent

The Shenzhen Intermediate People's Court accepted the bankruptcy petition of Yinghe (Shenzhen) Robotics in late August. The company had raised more than RMB 600 million in total and was once valued at US$500 million. According to information cited by Pencil Way (Chinese startup media), the company's 2025 revenue was only RMB 4.33 million, with a net loss of RMB 67 million. Li Lun of Panda Capital (Chinese VC firm), the angel-round lead investor, publicly criticized management over high salaries and defaulted orders, and Yingfeng Group has not responded publicly. Source: Sohu (Pencil Way) source

Momenta says its robotaxi fleet will reach "thousands" of vehicles by the end of next year · autonomy

Xie Shuo, head of the robotaxi business, said the fleet is now over 100 vehicles, will reach several hundred by year-end, and thousands by the end of next year. Munich and Abu Dhabi are under testing, with Dubai scheduled for next year. Cost reduction relies on in-house chips developed with XHeart. ⚠️ PlannedSource: Finimize source; Reuters source

UBTech plans its first overseas factory in Almaty · humanoid

Kazakhstan's President Tokayev met UBTech Vice President Zhong Yong and voiced support for building an education and service robot factory in Almaty. ⚠️ PlannedSource: Guandian Wang source