Today's Highlights

· Galbot robots now handle routine prescription-picking duty at a Laobaixing Pharmacy store in Shanghai, with the manufacturer claiming 400 to 500 fulfilled orders per day

· Agibot (Chinese humanoid robot maker) deployed 100 units of its Lingxi X2 across 100 Aishida (Chinese cookware brand) stores nationwide, greeting customers and recommending cookware

· Volvo's autonomous mining trucks have hauled a cumulative 3 million tonnes of material, triple the 2025 pilot target

· Momenta signed with Dongfeng Stellantis, putting world-model-based driver assistance into new Peugeot and Jeep models starting 2027

· Edge chip company SiMa.ai closed a $150 million Series C at a $1.45 billion valuation

· The world action model InternW0-Δ was pretrained on over 20,000 hours of open-source data, with code, weights, and data promised to be released

InternW0-Δ: A World Action Model Pretrained on Over 20,000 Hours of Heterogeneous Data · world-model

HF 17↑. World action models (WAMs — models that jointly predict visual changes and generate actions) have long lacked large-scale public training data. This paper aligns robot demonstrations, UMI, first-person human demonstrations, and Ego2Robot data into a unified state-action representation, assembling over 20,000 hours of corpus data, which the authors say is the largest open-source corpus of its kind. Architecturally, a Mixture-of-Transformers lets a video expert and an action expert interact under semantic guidance from a frozen VLM, with training-time distillation injecting 4D geometric priors so that future video rollout isn't needed at inference. It outperforms prior methods in both simulation and on real hardware, and the team says it will open-source the training code, weights, and processed data to the extent licensing allows.

Xingyu Miao et al. (InternRobotics) · arXiv 2609.31394 source

281-Gram Single-Leg Direct-Drive Robot Jumps 7.6 Meters High · locomotion

Without spring-based energy storage, a single long leg driven by a direct-drive motor jumps 7.6 meters, with a stance phase of just 190 milliseconds — the authors say this is the highest jump reported for an electrically driven system without elastic energy storage. The key is a fabric-wrapped transmission with variable mechanical advantage over the stroke, keeping the small motor near peak power throughout; a small propeller atop the leg handles attitude control both on the ground and in the air, and the system completed vertical and tilted jumps outdoors.

Gihyeok Na et al. · arXiv 2609.30530 source

Generate-Track-Refine: Ames' Team Fine-Tunes a Humanoid Locomotion Generator with Off-Policy RL · locomotion

Caltech's Aaron Ames group presents a two-layer architecture: a flow matching generator plans whole-body trajectories directly from depth images, while an underlying tracking policy handles execution. The contribution is an off-policy RL fine-tuning loop that is more sample-efficient than on-policy residual fine-tuning, improving terrain traversal success rate by up to 25 percentage points and skill selection by up to 80 percentage points. The same policy pair let a Unitree G1 walk, run, jump onto boxes, and climb stairs outdoors without relying on odometry or a height map.

Zachary Olkin et al. (Caltech) · arXiv 2609.31577 source

ECO: Leaving the Policy Untouched and Only Fixing Intermediate Waypoints Doubles Closed-Loop Driving Scores · autonomy

The authors find that the open-loop/closed-loop gap in end-to-end driving policies is concentrated in intermediate waypoints, while endpoint predictions are relatively reliable. They add a post-processing layer that anchors the executed history, preserves the endpoint, and reshapes the intermediate trajectory — with no need for maps, privileged state, or additional training. Closed-loop scores improved across all six tested policies, with VaVAM rising from 18.1 to 31.0 HD-Score on HUGSIM, taking first place in the HUGSIM closed-loop driving challenge.

Brayden Zhang et al. · arXiv 2609.31383 source

When Hierarchical VLA Planners Are Too Slow and Executors Don't Listen: Cutting VLM Forward Passes from 57 to 8 · vla

Using a diagnostic pipeline with a waypoint hierarchy adapted from π0.5, the authors find that erasing the waypoint endpoint has almost no effect on success rate, indicating the executor barely uses the planner's output. After adding block-autoregressive decoding with normalized goal modulation, VLM forward passes on LIBERO dropped from as many as 57 to 8, real-robot planning latency on a Rokae dual-arm setup fell 8.7×, and LIBERO-Long success rate rose from 91.0% to 96.2%; success rates on three dual-arm real-robot tasks matched the baseline.

Chuanliang Xie et al. · arXiv 2609.30833 source

Zero-Shot Online Learning on an 11.5-Ton Hydraulic Excavator Matches in 20 Minutes What Previously Took Over 100 · manipulation

With no demonstrations and no simulation pretraining, a probabilistic dynamics ensemble model is learned directly on a real Menzi Muck M445 excavator, paired with sampling-based MPC. After 20 minutes of interaction, tracking accuracy matches that of previous learning-based controllers that required 100 to 150 minutes of data; after 40 minutes, it maintains sub-centimeter average path error during high-speed operation.

Claudio Canales et al. (ETH Zurich et al.) · arXiv 2609.31025 source

VLaRL: Residual RL Trained Entirely in Simulation Still Improves a Frozen VLA on Real Hardware · vla

Since real-robot RL is expensive and risky, this paper treats the VLA's internal vision-language latent representation as a sim2real interface: a residual policy is conditioned on the latent representation, and a lightweight mapper is trained to pull simulated latents toward the real distribution. Across four contact-rich tasks and two VLA backbones, real-robot success rate improved for every combination, with no real-robot RL or online adaptation required.

Namiko Saito et al. · arXiv 2609.30868 source

SAIL: Adding Test-Time Scaling to VLM-Generated Robot Trajectories · manipulation

A collaboration between Sakana AI and the University of Tokyo, accepted at IROS 2026. A policy VLM generates trajectories from a handful of successful demonstrations, first trials them in simulation, then an evaluator VLM watches the replay video to pinpoint where things went wrong, and MCTS search proposes revisions — only the selected trajectory is sent to the real robot. Across six simulated manipulation tasks, as the number of candidates increased from 1 to 45, the average rate of finding a successful trajectory rose from 25% to 73%; the team also validated on real hardware, though it acknowledges how much of the improvement transfers to real robots remains an open question.

Sakana AI × University of Tokyo · IROS 2026 · Source: Sakana AI source

SiMa.ai | Series C | $150 Million | $1.45 Billion Valuation · hardware

Fidelity and Amplify co-led, with Dell Technologies Capital, StepStone, and others participating; AllianceBernstein, Baron Capital, and J.P. Morgan are new investors. Per the Business Wire release, the company has raised a cumulative $500 million. SiMa.ai builds embedded physical AI chips and accompanying development environments; the new funds go toward its next-generation Modalix SoC, targeting delivery in the first half of 2028 of hardware with 1,000 TOPS of compute. For reference, SiliconANGLE notes Nvidia's Jetson AGX Thor delivers 2,000 FP4 TOPS, while the mainstream T2000 and T3000 modules deliver roughly 400 and 865 TOPS. SiMa.ai is positioning on low power consumption, targeting mid-to-high-end drones, humanoid robots, and automotive ADAS.Source: SiliconANGLE source; Business Wire source

DEEP Robotics | STAR Market IPO | RMB 2.503 Billion Target | Updated Prospectus · embodied

The updated prospectus projects January–September 2026 revenue growth of 97.52% to 117.76% year-on-year, with median gross profit of about RMB 192 million, up from RMB 106 million in the same period last year; non-recurring-adjusted net profit attributable to shareholders is negative and declined year-on-year, which the company attributes to increased R&D spending, high-volume delivery of lower-margin models, and rapid team expansion. Revenue is still dominated by quadrupeds: 2025 revenue was about RMB 337 million, with the Jueying X series contributing about RMB 196 million, while the DR series of humanoid robots sold just 1 unit for the full year. Ifeng reports the company is currently in the inquiry stage of the IPO process, with CITIC Securities as sponsor.Source: CNFOL source; Ifeng Finance source; Ifeng Finance source

Xinhe Robotics (Shenzhen) | Seed Round | Tens of Millions of RMB · industrial

The investor is Shenzhen Gaoxinke Investment, a Shenzhen municipal state-owned investment arm. The company specializes in intelligent grain-leveling robots for grain storage facilities, working with the Chengdu Grain Storage Research Institute (affiliated with Sinograin) to advance deployment; the funds will go toward replicating the system at grain depots nationwide and validating new use cases such as ship cargo holds.Source: DoNews source

Mantis Robotics (US) | Strategic Investment | Amount Undisclosed · industrial

Japanese chemical company DIC has taken a stake in Mantis Robotics, which builds cageless high-speed industrial robots.Source: marketscreener source

Galaxea (Chinese embodied AI startup) Robots Now on Routine Duty at a Laobaixing Pharmacy Store in Shanghai · embodied ⚠️ Manufacturer's figures

At an instant-retail warehouse behind the Xuhui District Hongcao Road store, a Galbot G1 autonomously accepts orders, identifies and picks medications from shelves, verifies and packs them into a smart pickup locker, and delivery riders enter the last digits of their phone number to retrieve orders. Per the manufacturer's disclosure, this store currently fulfills 400 to 500 orders per day, with an average per-order handling time of 68 seconds and a picking accuracy of 100%, covering tens of thousands of product SKUs; nighttime operation capability is still being refined, and round-the-clock operation remains in the exploratory stage. Galaxea previously obtained a pharmaceutical retail operating license for humanoid robots and deployed one at a Beijing Neptunus pharmacy, where the company says labor input at that store fell by more than half. Xie Zilong, chairman of Laobaixing Pharmacy, said that with robots handling routine workflows, pharmacists can focus on symptom assessment and medication guidance. All operational figures above come from manufacturer materials, with no independent data yet available.Source: Sohu source; Gasgoo source

Agibot Deploys 100 Lingxi X2 Robots as Store Greeters Across Aishida Stores Nationwide · humanoid

Starting September 28, 100 units of the Lingxi X2 are being rolled out across 100 Aishida (Chinese cookware brand) stores nationwide, tasked with greeting customers, asking about their needs, explaining products, and recommending items. The first batch is doing three-day in-store livestreams across three cities — Wenling, Shanghai, and Wuxi — working 10 hours a day. A Wenhui Daily reporter watching the livestream observed the robot asking about household size, cooking habits, and budget before giving a cookware recommendation in about 30 seconds; when customers asked about specific discounts, it handed off to a human sales associate nearby to close the sale. The rollout doubles as a marketing livestream campaign, and whether stores continue using the robots routinely afterward remains to be seen once the livestreams conclude.Source: Sina Finance (via Wenhui Daily) source; Sina Finance source; Eastmoney source

Volvo's Autonomous Mining Trucks Have Hauled 3 Million Tonnes Cumulatively · autonomy

Volvo Autonomous Solutions announced last week that its fleet of autonomous heavy mining trucks has surpassed 3 million tonnes of cumulative autonomous material haulage, with Electrek reporting the milestone was reached at the Boliden Garpenberg mine in Sweden. The solution began piloting in 2023 at two customer sites — Garpenberg and Norway's Brønnøy Kalk — with an initial target of hauling 1 million tonnes by 2025. VAS Chief Technology Officer Ingo Stürmer said the figure "demonstrates the maturity of the autonomous hauling solution." Volvo sells vehicles, software, on-site infrastructure, and operational support as a bundled service called "Autona / earth," letting mines avoid building their own autonomous-driving teams.Source: Electrek source

Momenta Enters New Peugeot and Jeep Models, Mass Production Starting 2027 · autonomy

On September 24, Momenta signed a global strategic cooperation agreement in Shanghai with Dongfeng Stellantis Automotive Technology (Wuhan) to develop driver assistance based on its R7 world model for new Peugeot and Jeep models to be sold in markets including China and Europe. Momenta will participate in product definition, jointly tuning steering, braking, and power response to match Jeep's off-road character and Peugeot's handling characteristics. Per the automaker's statement, multiple new electric vehicles equipped with the system will launch progressively starting in 2027, with the cooperation also covering autonomous driving including L3. CEO Cao Xudong said the company is preparing for what it sees as an inflection point in the global driver-assistance market by 2028. Per its interim report, the company added about 321,000 vehicle installations in the first half of the year, up 83.7% year-on-year, with cumulative installations exceeding 1 million.Source: CnEVPost source

Aurora Sets 2030 Target of 30,000 Autonomous Trucks; Top Carrier Customer Says "There's Still a Gap in the Economics" · autonomy ⚠️ Company projections

At an investor day on September 23, Aurora set a 2030 target of operating more than 30,000 autonomous trucks, with revenue exceeding $5 billion and gross margin around 60%. Near-term milestones: 200 trucks by the end of 2026, corresponding to $80 million in annualized revenue; more than 1,000 by the end of 2027. The gross-margin breakeven point, previously expected around 500 trucks by the end of 2026, has slipped to roughly the first half of 2027. CFO David Maday called this "just a somewhat slower fleet ramp." The 70% gross margin target for 2028 set in 2024 will "likely be a year late," he said. Daragh Mahon, Chief Information Officer of carrier partner Werner, said at the same event that the two companies are still negotiating contract and economic terms, saying "frankly there's a gap between us," and that the numbers will only truly work out once the operation reaches scale, with neither side subsidizing costs. The company says it has now accumulated more than 500,000 autonomous miles since commercial operations began.Source: FreightWaves source

The Information: Optimus Production Reaches Hundreds of Units Weekly in August; Employees Resist Wearing Motion-Capture Suits to Train the Robots · humanoid ⚠️ Unverified report

Following Tesla's audit of Chinese suppliers in mid-September, ForkLog cites a report from The Information saying Optimus production reached hundreds of units per week in August, up from just tens of units per week in Q2; management's goal is to build a continuous automated production line with weekly capacity exceeding 1,000 units by year-end. The bottleneck lies in design: the arms and forearms have more than 100 screws and small parts still assembled by hand, and imprecise part fit means more units require rework after final assembly; hand sensors also have reliability issues, with plans to switch to replaceable sensor "gloves" next year. The Texas and California factories previously required employees to wear motion-capture suits at work, but employees resisted once they realized they were training robots to replace themselves, prompting the company to shift to using a dedicated team for data collection along with added cameras on the production line.Source: ForkLog source

Sharpa (Chinese embodied AI startup) Unveils Humanoid Platform, Dexterous Hand, and Data Glove at IROS · embodied

At IROS in Pittsburgh, Sharpa unveiled its integrated tactile dexterous manipulation robot, the D01: the arm has a load-to-weight ratio near 1:1, end-effector top speed exceeding 10.5 m/s, repeat positioning accuracy of 0.2 mm, and body-covering electronic skin. Its next-generation dexterous hand, the W02, has 21 active degrees of freedom, is about 30% smaller by volume than the W01, and its fingertip visuotactile sensors have a force-sensing range of 5 mN to 30 N. The exoskeleton data glove AE01 uses 22 encoders to capture hand motion, providing 256 levels of tactile feedback per fingertip, for teleoperation and data collection. Live demonstrations included catching a falling stick, teleoperated cocktail mixing, and real-time mirrored toy retrieval.Source: LeiPhone source; Sina Finance source

UBTECH Builds First Overseas Factory in Kazakhstan; Controversy Over U1 Deliveries · humanoid ⚠️ Company projections

Yerlan Nabiyev, CEO of Kazakh partner Nero Group, told The Astana Times that initial investment in the localized production project with UBTECH is expected to approach $440 million, covering R&D, programming, and manufacturing. Kazakhstan's investment agency says the Almaty factory for humanoid and educational robots is planned to begin production in Q1 2027, marking UBTECH's first manufacturing facility outside China; educational robot production is set to start in October, supplying nearly 30 Almaty schools by year-end. In China, after UBTECH announced on September 17 that U1 deliveries had begun, reports emerged saying U1 Pro deliveries had been pushed back to November 30. UBTECH responded that the reports were inaccurate and that it is shipping in batches according to pre-order sequence; on the same day, an Observer reporter contacting UBTECH's JD.com flagship store was told by customer service that "it's currently pre-order only, not yet delivered."Source: Sohu (via Central Asia Industry Research) source; Guandian source

China Holds 60.8% of Humanoid Robot Patents Across the World's Five Major Patent Offices · adjacent

Data obtained by the office of South Korean lawmaker Lee Jung-heon from the Korean Intellectual Property Office shows that, as of the first half of 2026, China accounted for 60.8% of humanoid robot patents among the IP5 patent offices, up from 51.8% in 2024; the United States fell from 27.3% to 21.8%, and South Korea fell from 6.3% to 4.6%. In first-half-year filings, China had 1,584 applications, the US had 113, and South Korea had 35; the Korean side notes patent disclosure has an 18-month lag, so recent figures remain provisional. Among robot foundation model patents, China's share reaches 66.3%.Source: The Korea Herald source

Mobilint Partners with LG, Doosan, and Daedong on Edge NPUs for Robotics · hardware

South Korean AI chip company Mobilint has joined four projects under the Ministry of Trade, Industry and Energy's "K-Edge AI Semiconductor" program: developing an edge accelerator for LG Electronics' humanoid robots, developing a chip and SDK for Doosan Robotics' next-generation collaborative arms, developing a chip and development environment for Daedong's unmanned agricultural robots, and joint R&D participation in an LG AI home appliance project. The approach incorporates customer product requirements from early in development and validates performance and cost on real hardware.Source: CHOSUNBIZ source

Yichang Unveils Embodied Robot Built on a Fully Chinese-Made Electronic Architecture · embodied ⚠️ Sole-source claim

Unveiled September 28 in Yichang, Hubei, and described by its developer as the world's first of its kind. Per a CCTV report, the National Industrial Information Security Development Research Center tested it under conditions including joint damage, cyberattacks, and virus injection; robots with conventional architectures experienced circuit failure or loss of system control, while this robot maintained safe operation. The report did not disclose the manufacturer or specific chip models involved.Source: Sina News (via CCTV) source

Anhui Issues Implementation Guidance for Its Robotics Industry, Naming Gaps in Reducers, Dexterous Hands, and Other Components · adjacent

A document from the General Office of the Anhui Provincial Government calls for consolidating the industrial robot sector while prioritizing humanoid robots, with focus areas including emotional interaction, commercial services, and elderly care rehabilitation; for core components, it prioritizes high-precision sensors, reducers, motors, batteries, dexterous hands, and joint modules, while also supporting aftermarket services such as repair, leasing, and recycling/disassembly.Source: Sina Finance (via Gelonghui) source