Today's Highlights

· IFR breaks out humanoid robots as a separate category for the first time: about 7,000 units sold globally in 2025, most going to research institutions and data collection

· Hai Robotics (Chinese warehouse robotics maker) will deploy over 1,500 shelf-climbing robots for a European fashion retailer

· Einride is rebuilding its next-generation autonomous driving system on NVIDIA Drive Hyperion

· Zhiyuan Robotics' (Chinese humanoid startup) Qiyuan unit launched two personal robots starting at RMB 19,999, shipping from October 1

· NHTSA and SAE launch the ASCEND coalition to develop federal performance standards for autonomous driving

· PARTS, a real-robot subtask RL framework, lifts Franka long-horizon task success rate from 50% to 95%

PARTS: Real-robot RL only on the bottleneck subtasks · manipulation

Base policies running long-horizon tasks tend to fail repeatedly on just a handful of subtasks, making it inefficient to re-collect demonstrations for the whole task. PARTS freezes the pretrained policy and trains a residual correction only on the bottleneck subtasks, with an agent-generated selector and success classifier providing local rewards. On the dual-arm YAM, full-task success rose from 32% to 61%; on single-arm Franka, from 50% to 95%, using only tens of minutes of real-robot rollout per task on average — more than 25 percentage points above existing real-robot RL fine-tuning methods at the same budget.

Sichang Su et al. · arXiv 2609.21788 source

VLA-Feedback: Diffusion VLAs that watch the latest frame even while executing an action chunk · vla

Action chunks generated by diffusion-based VLAs are usually executed open-loop, so they fall behind as soon as an object moves. This paper repurposes the last denoising step as a lightweight feedback interface: diffusion planning runs at low frequency, while a high-frequency loop corrects actions one at a time using the latest observation, without rerunning the full VLM. On static LIBERO tasks it matches GR00T; on dynamic simulated tasks, average success rate rose from 27.5% to 85.0%, and on real robots from 51% to 73%.

Yiheng Ji et al. (incl. Luis Sentis) · arXiv 2609.21022 source

AtomEgo: How to mix first-person human data into embodied pretraining · vla

Supplementing robot data with first-person human video is a bet several companies are currently making, but a systematic comparison of how to mix the two has been lacking. AtomEgo compiles roughly 2,659 hours of corpus and compares joint training, progressive transfer, and joint video-action modeling across both VLA and world-action-model architectures, backed by multi-task real-robot experiments. The takeaway in one line: gains scale with data volume times alignment quality — without good alignment, more data won't buy generalization.

Di Wu et al. · arXiv 2609.21461 source

GALA: Adding finger-level 3D geometry to latent actions · vla

Image-based latent action models can't capture fine finger-level motion, making it hard to combine human hand video with dexterous-hand data. GALA introduces a unified end-effector motion representation (UEMR) that combines visual latent actions with 3D end-effector geometric latent actions, providing supervision for cross-embodiment VLA pretraining — including unlabeled human video. It reaches 68.3% success on RoboCasa-GR1 and 75.5% on real robots; code is public.

Yichen Liu et al. (incl. Jianyu Chen) · arXiv 2609.21948 source

MT-WAM: A world-action model that gets stronger without generating video · world-model

Building on Fast-WAM's approach of not generating future video at inference time, this paper adds two extra supervision signals during training — predicting future 2D point trajectories and visual features — so the video backbone learns more of the dynamics information control actually needs. With no extra embodied pretraining, it reaches 98.2% success on LIBERO and 73.7% on LIBERO-Plus, 23.8 percentage points above Fast-WAM; average success on four real-robot tasks rose from 67.0% to 77.8%.

Yiguang Yang et al. · arXiv 2609.21474 source

ForceTwin: A hand-held force-sensing gripper puts a door's "feel" into its digital twin · perception

Two doors that look identical can feel completely different to push open, yet most existing digital twins only reproduce kinematics. ForceTwin has a person operate objects with a hand-held force-sensing gripper to identify inertia, friction, and damping, using a neural residual to capture state-dependent mechanism forces like those from a door closer. Feeding the identified results as a feedforward dynamics model into impedance control on Spot and Franka FR3, the target completion rate across nine object-embodiment combinations was 87%, versus 60% for a VLM-prior twin and 57% for a purely kinematic twin.

Tim Engelbracht et al. (incl. Marco Hutter, Marc Pollefeys) · arXiv 2609.21751 source

SkelWAM: A skeleton representation for zero-shot transfer of Franka-learned skills to other robots · manipulation

SkelWAM builds a 25-dimensional shared state from the arm's centerline, TCP pose, and gripper commands, which per-embodiment constraint decoders translate into joint commands — no demonstrations or policy updates needed on the target robot. The authors also release a cross-embodiment benchmark, LIBERO-Cross10, covering 10 target embodiments. Trained only on Franka data, it achieves 43.3% success on this benchmark, 36.2 percentage points above the strongest baseline, with additional real-robot demonstrations on a continuum robot.

Pengjun Niu et al. (incl. Hang Zhao) · arXiv 2609.21983 source

LIMBO: A 29-DOF humanoid learns to dodge balls and duck under bars, deployed on hardware without an online safety filter · locomotion

LIMBO learns a state-action control barrier function from black-box transition data, then distills that safety structure into the task policy so no online safety filter is needed at deployment. On a 29-DOF humanoid it demonstrated dodging balls and ducking under low obstacles; by adjusting sampling concentration under the same safety constraint, it also learned a lean-back maneuver for passing under a bar.

Jake Gonzales et al. (incl. Aaron D. Ames) · arXiv 2609.22075 source

ZYT-World: A seven-camera driving world model compressed to single-step generation · world-model

Built for production closed-loop simulation, this model natively generates 4 fisheye views over 180° plus 3 pinhole views. The authors distill a 40-step bidirectional teacher model into a single-step streaming generator, speeding up generation itself by 107.7x while retaining over 90% of the teacher model's PSNR and SSIM. Evaluation used only an internal multi-view test set.

Boni Hu et al. · arXiv 2609.21712 source

PaXini (Chinese tactile-sensing startup) ｜ A-share IPO coaching filing ｜ over RMB 4 billion raised cumulatively ｜ post-money valuation over RMB 10 billion · hardware

On September 20, PaXini completed IPO coaching registration with the Beijing Securities Regulatory Bureau, with Guotai Haitong Securities as coaching institution. Its B++ round, filed in mid-September, newly discloses Samsung as an investor via a strategic stake. The company's core business is tactile sensing; it says its self-developed tactile sensor chips have shipped over 1 million units cumulatively. Industry shareholders include BYD, JD.com, TCL, and SAIC Motor, with BYD as the largest external investor; the two signed a data-collection cooperation agreement in August. Chief Strategy Officer Luo Xiaoheng said "most industry partners approached us on their own initiative, becoming customers first and then shareholders." Source: Securities Times source

SemiDrive (Chinese automotive-chip maker) ｜ Series C+, C++ ｜ roughly RMB 1 billion combined across the 2026 Series C · hardware

Investors include Waldencast Ventures — founded by Intel CEO Lip-Bu Tan — and the Shaanxi Investment Group fund, among others. SemiDrive originally made automotive-grade chips, having shipped over 14 million units across its full lineup, and was valued at RMB 14 billion after its Series B+ round in 2024. Part of this round's funding will go toward pushing automotive-grade products into embodied intelligence; the company already has three product lines — R1, D9, and E3-R — covering "brain, cerebellum, and joints," and says related chips are already in mass-production use at leading embodied AI companies. Whether automotive-grade mass-production capability can transfer directly to robotics still depends on orders and customer validation. Source: iFeng Tech (智东西) source

Zhongyi Embodied (Chinese robot fleet-management startup) ｜ Angel round ｜ nearly RMB 100 million · adjacent

Led by shareholders of Andachi Intelligent, with Shenzhen-Hong Kong Hengtian and Beijing Jungerui participating. The company builds unified access and dispatch for multi-brand robot fleets, focusing on tourism/culture, buildings, and inspection scenarios. Its Tiangong platform handles rental, custom development, and maintenance, and says it has connected over 40 robot manufacturers with close to RMB 100 million in maintenance and rental orders in hand. Source: PEdaily source

Fangqi Technology ｜ Angel round ｜ tens of millions of RMB · world-model

With participation from Tus-Star Ventures. Founded this year, the company's core team comes from Tsinghua University and focuses on a "semanticized world model" — placing embodied large models and world models into the same 3D semantic space. It says it placed second in Stanford's BEHAVIOR 2026 household-chores challenge and has signed strategic cooperation agreements with companies including Youdi Robotics (Chinese service-robot maker). Source: LeiphoneNet source

Hai Robotics wins European warehouse deal for over 1,500 HaiPick Climb units · industrial

A new e-commerce fulfillment center for a leading European fashion retailer will deploy over 1,500 HaiPick Climb shelf-climbing robots, all connected to a single system, using goods-to-person picking and double-deep storage. The retailer's name has not been disclosed. This retailer has already used HaiPick systems in multiple European warehouses; this is a repeat order. Hai Robotics calls it the largest shelf-climbing robot deployment globally; a delivery timeline has not been disclosed. Last week, in its updated Hong Kong IPO prospectus, Hai Robotics reported first-half revenue of RMB 1.118 billion. Source: Materials Handling World source

Kodiak's driverless trucks connect to roadside enforcement systems, starting in Texas and Louisiana · autonomy

Kodiak built an automated process with weigh-station bypass provider PrePass that pushes results of enhanced inspections on driverless trucks to enforcement officers. Officers then decide whether a truck can bypass a roadside checkpoint or issue other instructions. Under CVSA rules, driverless trucks must be confirmed free of safety defects by a certified inspector before departure, with each inspection valid for up to 24 hours. PrePass CTO Chas Wurster said "the next hurdle for scaling driverless trucks isn't inside the vehicle — it's the infrastructure outside it." Kodiak plans to begin driverless commercial operations on highways before the end of 2026. Source: Heavy Duty Trucking source

Torc completes full-speed driverless acceptance testing at a closed course in Texas · autonomy

Daimler Truck's Torc completed product acceptance testing on a multi-lane closed course with no safety driver on board, reaching speeds up to 65 mph (about 105 km/h) in a Freightliner Cascadia with redundant systems. CEO Peter Vaughan Schmidt said repeated driverless tests showed "impressive reliability." Testing remains confined to the closed course, one step short of driverless operation on public roads. The company targets commercialization in 2027. Source: Machine Maker source

Two dual-arm humanoids already packing cushioning material in a CJ Logistics warehouse · embodied

The two robots operate at the Olive Young distribution center in Yongin, Gyeonggi Province. On September 18, CJ Logistics also signed a memorandum with RLWRLD (Korean robot foundation model startup) to jointly develop foundation models for logistics robots: CJ will provide on-site data and define performance standards, while RLWRLD handles model development. CJ already made an equity investment in RLWRLD last year and plans to next expand applications to picking, sorting, and quality inspection. Source: Seoul Economic Daily source

Brilliance Auto rolls out its first embodied robot, launches a data factory in tandem · embodied

This "Quantum One" unit rolled off the line and was delivered in Shenyang, equipped with a robot "brain" from AgiBot (Zhiyuan Robotics' consumer brand); it has already been collecting data at Brilliance's seat and interior-trim plants, with a data collection and labeling team of over 200 people. This delivery is a single unit — closer to a starting point than mass production. Source: Sina Finance source

IFR breaks out humanoid robots as a separate category for the first time: about 7,000 units sold globally in 2025 · humanoid

This is the first time in IFR's (International Federation of Robotics) three decades of statistics that humanoid robots have been counted as their own category, covering industrial and professional-service use and excluding consumer and military categories. Secretary General Susanne Bieller told Reuters that humanoid robots remain "a small fraction" of the global robot total; many units sold last year went to research institutions or were used to collect data for AI models rather than performing production work. Automakers are early adopters, but factories "usually only have single digits, sometimes double digits." For comparison, global industrial robot installations totaled about 542,000 units in 2024, meaning 7,000 units is under 2% of that figure. Bank of America Global Research estimates 90,000 humanoid units will ship this year and 1.2 million by 2030; Guandian.cn attributed the 90,000 figure to IFR, while versions relaying the Reuters report note it is a Bank of America estimate. IFR's 2025 industrial and service robot data will be released September 24. Source: Asia TV News via Newswav source

Einride rebuilds its next-generation autonomous driving system on NVIDIA Hyperion · autonomy

Last week Einride and Lidl launched a cabless L4 truck in Germany; this week it announced that the next-generation Einride Driver will be built on NVIDIA Drive Hyperion, using NVIDIA's compute, sensors, and safety software, with model training on Blackwell and simulation scenarios generated via Cosmos. Volvo and Aurora's programs also use Hyperion. Einride currently operates several hundred electric trucks across the US, Europe, and the Middle East, and expects its fleet to reach 1,500-2,000 vehicles by 2028 based on existing customer demand, with roughly 80% suitable for automation. Its stock rose about 7% in pre-market trading after the announcement, but is still down more than 63% year-to-date. Source: TradingView source

Zhiyuan's Qiyuan unit launches Q1 and T1 personal robots starting at RMB 19,999 · humanoid

Qiyuan, a unit of Sanwei New Materials, launched these two products in Shanghai on September 20, with orders shipping from October 1. Q1 stands about 88 cm tall and weighs about 15 kg, with open-source shell structural parts; T1 can switch between a wheeled-legged humanoid form and a quadruped form. Both standard editions are priced at RMB 19,999, with the Q1 Explorer edition at RMB 26,999 and the T1 Pro at RMB 29,999. Unitree's R1, in the same price tier, currently starts at RMB 29,900. CEO Tian Hua said the production line is "shipping one unit every 2.5 minutes on average." Sanwei New Materials invested RMB 164 million in its robotics business in the first half of the year with no revenue yet, but has already received RMB 210 million in advance customer payments. The launch event gave no timeline for the household-chore model PrimeMind. ⚠️ Company claimSource: Times Finance source

NHTSA and SAE launch ASCEND to develop federal performance standards for autonomous driving · autonomy

The coalition, co-launched by NHTSA and SAE International, aims to use data to develop a set of performance standards for autonomous driving systems. Current FMVSS rules assume a conventional passenger vehicle, which doesn't fit vehicles without a steering wheel or pedals, such as the Cybercab or Zoox. NHTSA Chief Legal Officer Jonathan Morrison said these performance standards could eventually be written into the Federal Motor Vehicle Safety Standards, also reducing the problem of each state setting its own rules. No timeline or evaluation methodology has been announced yet. Source: Digital Today source

Quadruped robot shipments in H1: IDC and Counterpoint estimates differ by about 14,000 units · humanoid

IDC's data shows over 49,000 units shipped globally, up 92.5% year-over-year, with a market size over $400 million; Counterpoint's data shows nearly 35,000 units, with Unitree leading at 37% share. IDC's report also shows that education/research and home consumer use account for about 60% of shipment volume but less than a third of revenue; industry-grade applications such as government and industrial account for under 40% of shipments but take over two-thirds of revenue. Within industry-grade applications, DEEP Robotics (Chinese quadruped robot maker) ranks first. Source: IT Home via Sina source; Guandian.cn source

NVIDIA engineers to embed at LG Electronics' Seoul data factory next month · embodied

The two companies will set up a joint development center at the Data Factory in Yangjae, Seoul, where engineers from both sides will develop robot models on the same site. LG plans to increase the number of robots running on-site to several hundred by year-end, combining real and synthetic data to reach 100,000 hours for training its own robot foundation model. Source: Asia Business Daily source

Samsung sets up an AX/PI team within its chip division dedicated to fab robotics · industrial

The team sits within the DS division's Global Manufacturing & Infrastructure Technology unit, with equipment maintenance likely the first focus, taking over some repetitive work currently done by outsourced staff and on-site engineers. The project is still early-stage, currently collecting data and planning robot routes and parts storage locations. Source: SammyGuru source