Key Takeaways

· AMD acquires World Labs in an all-stock deal worth $8.2 billion, with Fei-Fei Li becoming AMD Executive Vice President and Chief Scientist

· Two firms' figures for global humanoid shipments in the first half don't match: IDC says nearly 25,000 units, Frost & Sullivan says 36,000

· The 10,000th humanoid force-control arm from Rokae (Chinese robot-arm maker) rolled off the line in Zoucheng, Shandong; adjusted net profit turned positive in the first half

· A single dexterous hand rotating Rubik's Cube layers: 99.0% success over 300 attempts, and all 10 2x2 cubes solved

· A humanoid robot plays badminton, rallying with a human over multiple shots on a real robot, including jumping to return the shuttlecock

· Knowin (Chinese embodied-AI startup) has closed 5 rounds in one year, raising over RMB 1 billion in total, with a JD.com-affiliated fund leading this round

Research Progress

Self-evolving coding agent: writing task state and execution as code · vla

VLAs and world action models map observations directly to actions, and fail when the scene layout or viewpoint changes even slightly. This paper, with a Hugging Face (HF) trending score of 87↑, brings the coding-agent approach to robots, writing task state, progress and failure recovery as code that can be inspected and modified. The system, HexaAnything, calls perception, planning and control tools, with the VLA as one of the tools; on RoboCasa365 it beats XR-1 VLA on unseen composed tasks and on overall success rate, and HexaModel, trained on its execution trajectories, outperforms the base model on every split.

Hongcheng Gao et al. · arXiv 2609.35432 source

After 123 rounds of autonomous skill discovery: where robot self-improvement gets stuck · manipulation

This is a negative-results report. The agent observed its own failures, wrote new skills and installed external models, running 123 rounds of improvement. Every change passed its tests, yet the target task "put the seasoning on the top shelf of the refrigerator" never succeeded once. The authors conclude the bottleneck is not the agent itself: the segmentation model recognizes the shelf but not "top layer"; long tasks always fail at the first step, so later skills are rarely practiced; and where the evaluator measures wrongly, the agent optimizes against it anyway.

Jiaming Wang (National University of Singapore) · arXiv 2609.31760 source

Recursive harness distillation: a large agent writes an operating manual for a small agent · manipulation

HF 25↑. A strong agent writes its experience intervening on a VLA into a playbook and hands it to a lightweight agent, then repeatedly revises it based on the lightweight agent's execution feedback, with no model parameter updates throughout. Real-robot success rises from 37.3% to 64.0%; on SimplerEnv Bridge, the lightweight agent with the playbook reaches 66.7%, versus 41.7% for a GR00T-only baseline.

Seungyeon Kim et al. · arXiv 2609.33378 source

FINGR: a single dexterous hand turning a Rubik's Cube on a real robot · manipulation

When turning a Rubik's Cube, some fingers must hold the cube while others push a layer to turn, release and reset, with frequent contact. FINGR represents the cube's geometry relative to each fingertip and has the model predict contact-force changes and layer-turn progress. On a real robot, 300 layer turns succeed 99.0% of the time, versus 79.7% for the base flow policy; combined with grasping and regrasping against the tabletop, all 10 scrambled 2x2 cubes were solved, with the full process averaging about 137 seconds.

Yutong Liang et al. · arXiv 2609.33973 source

Humanoid badminton: learning hitting skills from a small amount of human motion data · locomotion

The authors say this is the first humanoid racket-sport system on a real robot that rallies with a human over multiple shots and can jump to return the shuttlecock. The method has three layers: first expand sparse hitting motions into an executable skill space; then a high-level planner combines skills such as forehand and backhand online based on the incoming shuttlecock's state; finally an adversarial regularizer constrains how natural the motion looks.

Jingzhi Cui et al. · arXiv 2609.31840 source

Uni-VLaT: adding whole-body tactile sensing to a humanoid VLA · locomotion

When the contact area is occluded, vision and proprioception cannot tell what is happening. Uni-VLaT adds a whole-body distributed tactile pathway to the VLA and has it predict future tactile, proprioceptive and visual representations. Across five real-robot tasks the average success rate is 75%, 43 percentage points higher than a baseline without touch; on the "pat-back walking" task, it improves by 85 to 90 percentage points on both VLA backbones.

Zihao Wang et al. · arXiv 2609.35450 source

ZeroBot: one image to a 3D mesh, a manipulation policy trained in two minutes · manipulation

No human demonstrations, no policy pretraining, and no known object models needed. ZeroBot generates a complete mesh from a single view of the object with an image-to-3D model, then runs massively parallel RL in simulation. On a real robot, the average success rate across grasping, pushing, articulated objects and multi-stage tasks is 87%, with an average training time of 119 seconds.

Ivan Kapelyukh et al. · arXiv 2609.34010 source

DexRoam: collecting whole-body human demonstrations with a VR headset to train mobile bimanual dexterous manipulation · manipulation

Collection uses only a consumer VR headset plus a head-mounted stereo camera, with no external cameras or motion capture. With human demonstrations added to training, GR00T N1.7's average success rate rises from 29% to 56%, and pi0.5 from 32% to 57%; to match the performance of training on robot data alone, it needs half as many robot demonstrations.

Rui Zhou et al. · arXiv 2609.35761 source

RoboFoundry: evolving the whole system behind the agent as a policy · vla

HF 17↑. The framework turns execution trajectories into verified context and skill updates, rather than optimizing a single component. The authors report it lifts GPT-5.5 by 27.8% on EmbodiedBench, with Qwen3.7-Plus catching up to 70.3% against GPT-5.5's 72.7%.

Jingsong Liang et al. · arXiv 2609.32862 source

Other papers today: Evolving Dexterous Robots from Scratch (evolves free-form manipulators with no preset body structure, which are printed, assembled and run zero-shot on a real robot, arXiv 2609.33101 source); RLE-Bench (tests coding agents as robot-learning engineers across four workflows: control, policy learning, perception and estimation, and mechanical design, arXiv 2609.34210 source); Dexterous Tactile World Model (a tactile world model for dexterous hands, arXiv 2609.34286 source); RE-0 (verifiable recursive improvement for code-as-policy agents, arXiv 2609.32416 source); SEES (locates weak atomic skills from failures and adapts the VLA accordingly, arXiv 2609.32698 source); Find Something You Can't Do (agentic real-robot RL for VLA self-improvement, arXiv 2609.32069 source); WB-WAM (a world action model pretrained on heterogeneous humanoid whole-body data, arXiv 2609.34199 source); The Low-Rank Structure of VLA Reinforcement Learning (the low-rank structure of VLA RL updates, arXiv 2609.34599 source); X-Reset (cross-embodiment resets to scale object-centric RL, arXiv 2609.35715 source).

Open Source · Tools · Evaluation

· APXInf: an on-device inference engine for embodied models, open-sourced on September 15 by Tsinghua University together with Infinigence AI (Chinese AI compute startup) and Shanghai Jiao Tong University. Without changing the model, π0.5 inference latency on Thor FP8 drops from 278ms to 26ms, about 10.7x; LIBERO-10 success rate is 92.2%, versus 92.4% for the reference implementation source

· FLUX 3 Action: Black Forest Labs' 7B robot model (weights reported), which MIXED reports ranks first at 42.9% on Nvidia's own robot leaderboard ⚠️ Single-source claim source

Funding and Deals

World Labs | Acquired by AMD in all-stock deal | $8.2 billion · world-model

AMD is acquiring World Labs, the world-model company Fei-Fei Li founded in 2024, in an all-stock deal valued at $8.2 billion, and Li joins AMD as Executive Vice President and Chief Scientist. The two companies already formed an inference-optimization and training partnership last year. World Labs says AI R&D needs "close collaboration between model research, systems and compute," while AMD says frontier workloads will influence its chip roadmap. Li wrote in the announcement post that scaling up requires being "closer to the hardware." World Labs' first product, Marble, is used for entertainment content and also for generating simulation environments for robot training; Nvidia already has its Cosmos series of open-weight world models, while AMD had previously disclosed only text and video models. Sources: TechCrunch source; Huxiu source

Knowin | Angel+++ round | Several hundred million RMB · embodied

A JD.com-affiliated fund led, with Zhengxin Valley Capital, Nanshan Strategic Emerging Industries Fund and Hua Capital (Chinese VC firms) participating. Knowin was founded in August 2025 and has completed 5 rounds in one year, raising over RMB 1 billion in total, only 49 days after its previous RMB 500 million round announced on August 10. The company builds GLOW, an embodied foundation model for home settings, and the KNOWIN-X1 robot body, and plans to release and mass-produce in the first quarter of 2027. The company says GLOW averages 62.2% success on 18 RoboDojo simulation tasks, 40 percentage points above a GPT-6 baseline ⚠️ Vendor claim. Source: Pedaily (Chinese VC news outlet) source

Shanghai Shunxin Embodied Intelligence Robot Fund | Registration completed | RMB 2 billion · adjacent

The fund, managed by CSGC Shunjing (Chinese defense-industry fund manager), completed registration with the Asset Management Association of China on September 24, and mainly invests in unlisted companies along the humanoid robot industry chain; Gehua CATV (Chinese cable TV operator) subscribed RMB 100 million from its own funds. The same day, the Beiyang Embodied Intelligence Innovation Industry Fund announced its launch, also at RMB 2 billion. Sources: Pedaily source; Sina Finance source

IPercept | New round | $16.5 million · industrial

Builds factory equipment health monitoring. Source: Dealroom source

Company under Bluepoint Touch | Zhiyuan Robotics (Chinese humanoid startup) takes a stake | Amount undisclosed · hardware

Bluepoint Touch (Chinese robot force-sensing company) builds force sensing for robots. Source: 36Kr source

Commercialization and Deployment

Rokae's 10,000th humanoid force-control arm rolls off the line · embodied

The 10,000th humanoid force-control arm came off the line at the Zoucheng, Shandong plant on September 29. The company's announcement says these products run on the same production line with the same factory tests. According to the interim results announcement, first-half shipments of the full robot range exceeded 13,000 units, of which embodied-AI robots were 5,600 units; total revenue was RMB 415 million, up 135.8% year on year, with the embodied-AI business at RMB 139 million, or 34%. Adjusted net profit was RMB 18.4 million, turning positive from a loss a year earlier; consolidated gross margin was 29.8%. The Taibai AR series force-control arms have seven degrees of freedom and a 3 to 20kg payload. Source: Guandian (Chinese financial news outlet) source

Kodiak begins driverless produce hauling in California · autonomy

Driverless truck freight extends to fresh produce routes. Source: The Korea Daily source

Caterpillar deploys autonomous haulage system at Richards Spur quarry · autonomy

Cat's AHS gains another quarry site. Source: International Mining source

Yikong Zhijia (Chinese autonomous mining truck company) validates driverless mining trucks in a tunnel section at Zijinshan Mine · autonomy

This was validation and has not yet moved to routine operation. Source: International Mining source

German building-materials retailer Bauhaus uses XYZ Robotics for container unloading and palletizing · industrial

Used for container unloading and palletizing. Source: Robotics & Automation News source

UK courier Evri delivers parcels with a bipedal robot in one town · humanoid

The robot can climb stairs and hand parcels to recipients. The report covers only a single town, so this is a pilot. Source: The Sun source

Industry Developments

Global humanoid shipments in the first half: IDC says 25,000, Frost & Sullivan says 36,000 · humanoid

A tracking report released by IDC on September 29 says global humanoid robot shipments in the first half of 2026 were close to 25,000 units, up 432.1% year on year, and forecasts more than 750,000 by 2030. A report released the day before by Frost & Sullivan puts first-half sales at 36,000 units and revenue at $540 million; the two differ by about 11,000 units, and the reports do not explain where the gap comes from. Frost & Sullivan's data show a severe inversion between units and revenue: in the first half, small-size products sold 19,700 units but earned only $19 million, while full-size products kept a revenue share above 75%. Production-work scenarios sold only 1,200 units, with revenue of $78 million, up 65.9% on all of 2025. On vendor rankings, UBTech (Chinese humanoid maker) ranked first with first-half revenue of RMB 620 million and a 16.5% market share, but that is a clear drop from 24.8% in 2025; Zhiyuan follows at 16.1%. Sources: Guandian source; 21st Century Business Herald source; HK01 source

CNBC: China's securities regulator sets three thresholds for embodied-AI IPOs · humanoid

CNBC, citing three people familiar with the matter, says the China Securities Regulatory Commission (CSRC) has used window guidance to require embodied-AI companies seeking to list to meet three conditions: sustainable revenue and commercial orders; narrowing losses, with one source saying a three-year forecast is needed; and mastery of core technology such as a robot brain or dexterous hands. Another source said meeting two of them may be enough, but it remains unclear which companies can qualify. At least two dozen humanoid-related companies have filed in Hong Kong alone. Unitree (Chinese robot maker) closed at RMB 845 on its August 19 first day of trading, and as of this Monday it had fallen to RMB 459.65. ⚠️ Rumor-based claimSource: CNBC source

Deep Robotics prospectus: humanoid unit cost falls to RMB 147,100, with 6 units sold in six months · humanoid

Following the prospectus update on September 28, a breakdown by product line: the DR series humanoid sold 6 units in the first half at an average selling price of RMB 464,900 and a unit cost of RMB 147,100, both down more than 40% from 2025; it made up 1.08% of revenue. Carrying gross profit is the wheeled-legged Lynx M series, with 482 units in the first half and a gross margin of 60.04%. Net loss attributable to shareholders in the first half was RMB 8.8063 million. Founder Zhu Qiuguo says robots at this stage are still essentially "semi-automatic equipment." Sources: The Paper (Chinese news outlet) source; Securities Star source

Sharpa launches a humanoid body, dexterous hand and exoskeleton glove together at IROS · embodied

The new W02 dexterous hand drops from the previous generation's 22 active degrees of freedom to 21, removing the little finger's rarely used base rotation joint, with a 30% smaller volume and an IP54 protection rating; the AE01 glove on the data-collection side still keeps 22 degrees-of-freedom input. The D01 humanoid body extends tactile sensing to the arms and chest, and the vendor says end-effector maximum speed exceeds 10.5m/s ⚠️ Vendor claim. Sources: Sohu source; 36Kr source

Bolt wants 25,000 Lucid robotaxis, but the EU allows only 1,500 registrations per model per year · autonomy

Bolt announced on September 17 plans to deploy at least 25,000 Lucid driverless vehicles in Europe; the announcement gives no launch year, cities or financial terms. Under current EU type-approval rules, a driverless vehicle model can register only 1,500 units per year. Lucid produced only 4,774 vehicles in the second quarter. ⚠️ Plan-based claimSource: The Auto Wire source

Pony.ai's fourth-generation autonomous heavy truck unveiled · autonomy

The company says it will reach volume production this year. ⚠️ Plan-based claimSource: Autohome source

LG Innotek and Applied Intuition put an autonomous-driving test fleet into operation in South Korea · autonomy

The fleet is used to collect data for developing autonomous-driving sensing systems. Source: Taibo (Chinese tech news site) source

Tesla FSD approved in Croatia · autonomy

The EU-level FSD vote has been postponed to December at the earliest (reported); this is a clearance by a single member state. Source: Investing.com source

Hardware · Supply Chain

· Luoyang Bearing (LYC): some bearing products for industrial and embodied robots have entered customer validation and small-batch supply source

· EdgeCortix RAIDEN: a chiplet platform for physical AI applications has launched source

· BOS Semiconductors: joins Arm's Physical AI ecosystem, building custom customer chips source