Today's Highlights

· XPeng launched its first humanoid robot automated production line — IRON walks off the line on its own two legs, with mass production planned by year-end

· UBTECH says U1 has begun deliveries

· Fulongma paid Huawei Cloud RMB 200 million for exclusive development of an autonomous driving system for sanitation vehicles — about 20 times its unmanned sanitation revenue from last year

· D-Robotics (Chinese embodied-AI chip maker) completed a $400 million Series C, led by Mirae Asset with Meituan as a strategic investor

· Chelsea Finn's team applied RL fine-tuning to real-time VLA, raising average performance on four dynamic real-robot tasks from 42% to 97%

· Sharpa open-sourced a world synesthesia model, transferring pretrained priors to 49 new objects for in-hand dexterous rotation

RL fine-tuning for real-time VLA: the large model proposes action chunks, a small policy edits actions against the latest observations · vla

Large VLAs infer slowly, so by the time an action is executed the observation it was based on is already stale. Prior asynchronous execution schemes were built on imitation learning and couldn't be pushed toward higher reliability with RL. Chelsea Finn and Dorsa Sadigh's team split the two roles apart in EXPO-FT: a pretrained VLA slowly proposes action chunks, while a lightweight editing policy quickly corrects them against the latest observations. Across four real-robot tasks — handing over objects, balancing a ball, kicking a tabletop soccer ball, and grasping moving objects — with online data capped at 10 minutes and no human intervention throughout, average performance rose from 42% to 97%.

Perry Dong et al. · arXiv 2609.18207 source

World synesthesia model: compressing vision, touch, and action history into a reusable world state · manipulation

Accepted to CoRL 2026, code already open-sourced. In-hand dexterous manipulation tends to degrade into open-loop routines when it hits real depth-sensor noise. Sharpa uses a Dreamer-style RSSM to fuse proprioception, touch, and wrist depth, adding noise to inputs during training while using clean depth as the reconstruction target. On a 22-DoF five-fingered hand, priors pretrained on nine objects transferred to 49 new objects, averaging 9.37 rad/episode of rotation versus 3.28 rad without the prior; on a real robot, a cube could be rotated continuously and stably for over a minute.

Sharpa Robotics team · arXiv 2609.07002 source · Coverage: Sina Technology source

ActionPiece: an action tokenizer needs to preserve the relative distance between actions · vla

36↑ on HF. The authors point out that looking only at MSE hides a problem: after compression, the action adjustments needed in different contexts can be flattened or even reversed. They propose a "physical rank consistency" metric and use it to supervise representation learning and quantization. Under the same Qwen3-VL-4B training setup, this reaches 94.8% on LIBERO and 68.8% on the unseen LIBERO-Plus.

Shijie Lian et al. · arXiv 2609.18487 source

Diagnose before fine-tuning a VLA: which layers need adjustment can be measured in advance · vla

The authors found across five architectures — OpenVLA-OFT, π0, SmolVLA, DTP, and Octo — that adaptation cost for appearance changes concentrates in the vision encoder, while cost for instruction changes concentrates in the language backbone. Based on this, they estimate per-region cost using just 10 unlabeled target images, then assign LoRA of different ranks accordingly. On a real xArm-7 robot, when instruction phrasing changes, just 0.04% of trainable parameters matches the performance of full fine-tuning.

Shahram Najam Syed et al. · arXiv 2609.18084 source

Decoupling vision, language, and action: matching VLA performance without a VLM backbone · vla

DEM uses a fine-tuned DINOv3 as the vision encoder, a frozen NeoBERT as the language encoder, and a single-step MeanFlow head that outputs action chunks. Across 18 simulated tasks and 3 real-robot tasks, success rate matches SOTA policies with VLM backbones, while inference frequency is 8-17 times higher and per-inference energy is 6-15 times lower. The conclusions are limited to the authors' evaluation protocol and task scope.

Xiatao Sun et al. · arXiv 2609.18374 source

PointZero: filling in pretrained dynamics priors from 3D point trajectories, without robot action labels · world-model

Learning action-conditioned 3D dynamics usually requires robot action labels, which rules out using web video. PointZero instead predicts future 3D trajectories for all points from a single RGB-D frame and sparse trajectories, trained on 2.9 million frames of synthetic data. After post-training, it outperforms baselines on the PGND benchmark, matching or leading on 6 of 7 manipulation tasks. Dataset, weights, and training recipe are all released.

Bardienus P. Duisterhof et al. · arXiv 2609.19142 source

WholeBodyWAM: grounding a humanoid world-action model with 4,000 hours of human motion · world-model

Collecting whole-body trajectories for a target robot is expensive, while human and humanoid motion data are relatively abundant. The authors built UniMotion-4K, unifying human video, 3D motion libraries, and data from multiple humanoid platforms into a single action space; they first pretrain an action expert, then jointly post-train it with a video expert and the action expert via asymmetric MoT attention. Larger-scale action pretraining improves downstream performance, and it also transfers to real-robot humanoid manipulation.

Bowei Zhang et al. · arXiv 2609.18197 source

PASSAGE: 100 hours of scene-aligned human motion, teaching humanoids to step over, crawl under, and turn sideways through obstacles · locomotion

The team used VR and inertial motion capture to collect 100 hours of motion across 1,500 cluttered scenes, training a flow-matching planner plus a perception-aware whole-body tracker, with no skill annotation required. As the data scaled from 6 to 100 hours, the collision-free success rate on unseen scenes rose from 48.1% to 68.9%. The full system runs entirely onboard a Jetson AGX Orin and succeeded across 50 real-world layouts with no pre-built map.

Yuxuan Ma et al. · arXiv 2609.18732 source

KINO: connecting VLM planning to humanoid whole-body control via action keyframes · locomotion

A VLM selects whole-body pose keyframes from a preset library, retargets them to object poses, and an RL policy handles reaching those targets. The key is saliency-based keyframe sampling: with sparse keyframes, end-to-end success rate rose from 44% to 92%. The system performed single- and dual-handed picking, carrying, and placing on a Unitree G1.

Sitong Chen et al. · arXiv 2609.18869 source

GPT-Policy: letting a general-purpose VLM agent learn robot skills in-context from demonstrations · manipulation

13↑ on HF. The framework consists of a context compiler, a VLM that proposes tool actions (such as GPT-6 Astra), and a constrained controller responsible for verifying execution — no gradient updates involved. In real-robot trials, human video demonstrations without action labels still improved completion rate, and aligned action references gave a further boost on contact-sensitive tasks.

Dongzhou Cheng et al. · arXiv 2609.19138 source

D-Robotics ｜ Series C ｜ $400 million · hardware

Led by Mirae Asset, with Meituan as strategic investor; Hefei Guotou, Nanshan Strategic Emerging Industry Fund, Cathay Capital, and GF Xinde also participated, alongside follow-on investment from existing shareholders Hillhouse Ventures, 5Y Capital, and Vertex Growth. D-Robotics was incubated by Horizon Robotics (Chinese autonomous-driving chip maker) and builds robotics compute chips plus a development platform. The company says first-half revenue grew several-fold year-on-year, with cumulative shipments of its Sunrise-series chips passing 8 million units. Six months after the Sunrise S600's launch, more than 20 customers including AgiBot, Galbot, X Square Robot, and UBTECH have adopted it. New funds will go toward the full compute-range Sunrise chip lineup and a software platform spanning data collection through inference deployment.Source: PEdaily source

Emulate ｜ Seed round in talks ｜ Up to $700 million ｜ Pre-money $3-3.7 billion · world-model ⚠️ Reported, unconfirmed

The company was founded just this past August; founders Jack Parker-Holder, Matthew McGill, and Philip Ball all came from Google DeepMind's world-models team, with Parker-Holder having been a lead researcher on Genie 3. Index Ventures and Lightspeed are leading, with Creandum participating. The two outlets' valuation figures differ: FT reported a pre-money valuation of $3.7 billion, while Bloomberg reported the company seeking a pre-money valuation of $3 billion; FT said terms had not yet been finalized.Source: Tech Funding News source

Viabot ｜ Series A ｜ $24 million · adjacent

Led by Walden International, with participation from CDIB Capital and Stalwart Ventures, bringing total funding to $43 million. Viabot One cleans debris and performs patrols in large commercial properties; the company says it currently serves about 25 million square feet of property (roughly 2.32 million square meters), with customers including Fortune 500 retailers and universities.Source: Pulse 2.0 source

XPeng's humanoid robot automated production line goes live, IRON walks off the line · humanoid

At the G9L launch event on the evening of September 17, He Xiaopeng announced the line was operational, which XPeng calls the world's first automated production line for advanced general-purpose humanoid robots. The line has no hoisting or conveyor belts — IRON walks off the line on its own two legs at the end. XPeng says the entire process, from component assembly to final vehicle-style assembly, is fully automated. IRON stands 178 cm tall, has 76 degrees of freedom across its body, and carries 3 Turing chips with total compute of 2,250 TOPS, capable of running a 30-billion-parameter model locally. Under XPeng's plan, mass production begins by year-end, with a monthly capacity target of over 1,000 units, initially deployed at XPeng stores and campuses, with sales and delivery in China and overseas markets starting in 2027. The production line going live is confirmed; the monthly capacity figure remains a target that can only be verified once the ramp-up happens by year-end.Source: Kuaikeji source

UBTECH's U1 hyper-bionic humanoid begins deliveries · humanoid

On September 16, UBTECH CEO Zhou Jian announced that deliveries of the U1 series had begun, with the first recipients being Shanghai Yongda Group, Beijing Boshi Group, and South Korea's Galaxy Robot Land; Boshi will roll out U1 units in stages into experience spaces at its Mercedes-Benz and other brand dealerships. U1 has 88 degrees of freedom; the Lite version costs RMB 119,800, the Pro version RMB 169,800, and the Ultra version RMB 880,000-990,000. UBTECH previously said orders across all channels had exceeded 13,000 units since the launch in late June, but did not disclose delivery volume this time.⚠️ Company-reported figuresSource: Gasgoo source

AgiBot unveils commercial use cases for Yuanzheng A3 Ultra · humanoid ⚠️ Company-reported figures

Use cases span dealerships, hotels, and convenience stores, with tasks including guest greeting, room tidying, night patrols, and shelf restocking. The robot carries a 700 TOPS compute platform and supports autonomous charging and battery swapping. AgiBot says it has completed "thousand-unit-scale mass production validation" but did not give deployment figures for each scenario.Source: Guandian.cn source

Fulongma signs RMB 200 million autonomous driving development contract with Huawei Cloud · autonomy

Total contract value is RMB 200 million (tax included), over a 3-year term. Huawei Cloud will exclusively develop and deploy an autonomous driving system for Fulongma's sanitation vehicles, covering 3-ton and 6-ton class vehicles and both sweeping and wash-sweeping operations. Payment is split into three milestone-based installments of 15%, 75%, and 10%. Models trained on augmented data from Fulongma belong to Huawei Cloud, and Fulongma must subscribe to Huawei Cloud services to use them. The filing discloses that Fulongma's unmanned sanitation product line generated RMB 9.9097 million in sales revenue in 2025 and RMB 4.9513 million in the first half of this year. The contract value is about 20 times last year's revenue from this segment; the company also flagged uncertain impact on profit.Source: Guandian.cn source

Lucid and Bolt sign European robotaxi partnership, targeting at least 25,000 vehicles · autonomy ⚠️ Planned, not yet executed

Bolt plans to deploy at least 25,000 fully self-driving vehicles based on Lucid's not-yet-in-production midsize vehicle platform, with the fleet owned and operated by Bolt itself using Nvidia's Hyperion compute and sensor architecture; the autonomous driving software supplier has not been disclosed. According to TechCrunch, no money has changed hands between the two companies, Bolt has not placed an order, and there is no deployment timeline; the first vehicle on this midsize platform was just delayed by nearly a year, to the second half of 2027. Bolt founder and CEO Markus Villig said autonomous driving in Europe needs "data, software, vehicles, and operations working together as one system, built for European roads and regulations."Source: TechCrunch source

Volkswagen's MOIA fails to secure outside investors as Uber and Lyft exit talks · autonomy

Uber and Lyft were the final two bidders, and both have abandoned plans to take a stake in Volkswagen's robotaxi division. Disagreements centered on three points: billing per kilometer versus per vehicle, which vehicle model to use, and contract duration. Another sticking point was whether Volkswagen itself would keep investing further. MOIA is now funded solely by Volkswagen.Source: electrive source

Waymo partners with Allianz on European robotaxi insurance, starting in Germany · autonomy

Allianz Partners will provide fleet insurance, multi-country product liability coverage, and digital claims processing for Waymo's fleet, with the Allianz Technology Center involved in accident analysis and safety research. Waymo plans phased commercial deployment starting in Germany. Waymo's head of risk and insurance, Tilia Gode, said the partnership will let the company "expand reliably and responsibly from day one."Source: Life Insurance International source

XPeng upgrades second-generation VLA, edge-side parameter count up 3.5x · autonomy ⚠️ Company-reported figures

He Xiaopeng said the new version incorporates "time" into the physical-world foundation model, supporting up to 30 seconds of effective temporal reasoning. The edge-side model's parameter count is more than 15 times that of mainstream VLAs. XPeng also claims end-to-end response speed improved 300% and multi-dimensional comprehensive safety capability improved 20-fold, though it did not disclose the testing methodology.Source: Sina Finance source

XPeng plans to sell technology to foreign automakers beyond Volkswagen · autonomy ⚠️ Reported, unconfirmed

Reuters cited two people familiar with the matter as saying XPeng intends to offer its electronic and electrical architecture, cockpit systems, Turing AI chips, and driver-assistance software externally, and has already been in contact with some interested automakers. Its first joint vehicle with Volkswagen, the ID.UNYX 08, went into mass production in March this year.Source: Reuters source

Jiushi builds its own compute cluster of nearly 15,000 GPUs · autonomy ⚠️ Company-reported figures

Jiushi (Chinese autonomous delivery vehicle maker) says this makes it the first L4 company to build a 10,000-GPU-class cluster. The cluster is used to train the APEX world model (starting at tens of billions of parameters), with training data including 270 million kilometers of L4 operational mileage. The onboard system is tiered by vehicle speed: above 30 km/h, an onboard end-to-end model drives independently; between 5-30 km/h, a cloud-based VLA is invoked alongside the onboard VLA; in narrow-lane or stuck-vehicle scenarios below 5 km/h, a "Safety Officer Agent" handles detours and replanning.Source: SmartHey source

Tutor Intelligence launches second-generation Cassie and Sonny warehouse robots · industrial

The single-arm Cassie has moved from a fixed workstation to being mobile, able to lift 50 pounds (about 23 kg) and drag loads of several thousand pounds. The dual-arm, semi-humanoid Sonny runs a 4.5-billion-parameter Ti0 VLA model, with data collected at the company's own Data Factory 1. CEO Josh Gruenstein said: "What matters to us isn't just building general-purpose robots, but making them useful right away."Source: SiliconANGLE source

Deep Blue Robotics enters bankruptcy liquidation, employee claims exceed RMB 23 million · adjacent

DoNews, citing StrongrTang, reported that the bankruptcy liquidation case for Deep Blue Robotics (Shanghai), accepted by the Shanghai Pudong court, completed its fifth round of employee-claims disclosure on September 7. Over 100 employees are involved in total, with combined employee claims exceeding RMB 23 million. The parent company had completed a pre-IPO funding round as recently as April 2025.Source: DoNews source

Samsung Heavy Industries signs agreement with Korea Institute for Robot Industry Advancement to jointly develop shipbuilding robots and industrial humanoids · industrial

The two sides signed an MoU on the 16th at the Geoje shipyard covering joint R&D and on-site testing, with no deployment timeline disclosed. The two have collaborated on welding robot development since 2015, and those robots are already in operational use.Source: Korea Startup Post source

Unitree's share price returns to RMB 500, total market cap back above RMB 200 billion · humanoid

After falling below RMB 500 on the 10th, Unitree rose 5.61% today to close at RMB 500.90.Source: Guandian.cn source

NHTSA orders Tesla to provide sworn response on Cybercab's compliance basis by September 30 · autonomy

Following the September 16 report, details of the special order have now been made public. Cybercab has no steering wheel, pedals, or mirrors, and Tesla has not applied for a Part 555 exemption. NHTSA is requiring Tesla to explain which federal safety standards it deems applicable or inapplicable, whether temporary manual controls could be added, and whether such controls would form the basis for self-certification. The response must be given under oath; an incomplete or false response carries a maximum penalty of about $139 million. NHTSA's 2022 rule states that vehicles operated solely by an automated driving system will "likely require" an exemption to be manufactured and sold before the relevant standards are revised.Source: Electrek

Minneapolis proposes legislation requiring a human in the driver's seat of robotaxis · autonomy

Four city council members introduced the "Workers Drive Minneapolis" ordinance: Waymo, Tesla, Zoox, and others would first need a city license, and every vehicle would need a licensed safety attendant in the driver's seat, able to directly control steering, braking, and acceleration, paid no less than minimum wage. Proponents say the goal is to protect the more than 10,000 rideshare drivers in the state; Waymo called it a "de facto ban." Even if passed, the ordinance would not take effect until October 2027, which proponents say is meant to give the state legislature time to develop statewide rules.Source: Minnesota Reformer source

Product launches put edge compute front and center, as robot chip funding and benchmarks land in the same week · hardware

Two humanoid robots this week both put compute figures into their headline specs: XPeng's IRON uses 3 self-developed Turing chips with total compute of 2,250 TOPS, while AgiBot's Yuanzheng A3 Ultra is rated at 700 TOPS. On the chip side, moves came from two directions. D-Robotics landed a $400 million Series C, saying that six months after the Sunrise S600's launch, more than 20 embodied-AI customers have adopted it, with customer coverage among embodied-AI firms surpassing 50%. Nvidia posted results on MLPerf's newly added edge agentic benchmark using Jetson AGX Thor. The two sides' evidence differs in kind: D-Robotics cites customer counts, while Nvidia cites a public benchmark score, and there is not yet comparative data on the same task. Upstream components showed weaker signals this week, with Luoyang Bearing saying its robot bearings remain in the customer validation and small-batch supply stage.Source: Kuaikeji source, PEdaily source, Guandian.cn source, NVIDIA Technical Blog