Today's Highlights

· Nvidia and Foxconn publish assembly data for GB300 test trays: robots install the busbar with a success rate above 95%

· Geek+ (Chinese warehouse robotics maker) has more than 1,600 robots working in two of LPP's fulfillment centers in Poland

· Einride's two cabless electric trucks haul freight on public roads in Ohio every day with no one on board

· Doosan Robotics leads two Korean national physical AI projects worth 98.9 billion won in total

· The hierarchical world model H-JEPA lifts Visual AntMaze success from 18% to 73%

· For robot instructions with implicit malicious intent, in up to 95% of episodes the task was completed and no text guardrail flagged it

H-JEPA: end-to-end training of a hierarchical world model for visual planning · world-model

A single-level JEPA world model predicts at only one time scale, so long-horizon planning tends to drift. H-JEPA lets each level predict further ahead in its own latent space, and the upper level's predictions serve as subgoals for planning at the level below. On Visual AntMaze, the three-level structure raises success from 18% to 73% while using less planning compute. The authors include Yann LeCun; the method was also applied to offline planning on real-robot DROID video.

Wancong Zhang et al. · arXiv 2610.06805 source

SUAVE: unifying video and action with masked diffusion · world-model

VLAs can see and act, but they do not imagine the future before acting. World action models built on video diffusion can imagine the future, but language is only a frozen conditioning input. SUAVE turns language, video and action into discrete tokens in one sequence and generates them with a single masked diffusion transformer; depending on which tokens are masked at inference, it acts as a world model, a policy, or a video-action model. On a real robot, it takes 1030 ms on an RTX 5090 to generate a subgoal image and an action chunk covering 1 second, giving a control rate of 2.5 actions per second.

Rhythm Syed et al. · arXiv 2610.04009 source

Mulligan: letting policy performance decide where data collection starts · manipulation

The further human demonstrations go, the smaller the gain from each additional one. Mulligan decides for itself which object placements each round of supervised deployment starts from, favoring initial states where the policy has failed or that it has not yet tried. It also trains a value function so that the failure data discarded by imitation learning can be used. Evaluated on 2,550 blind-test episodes across three high-precision real-robot tasks, final success rates improve by 10-34 percentage points. Code and data are public.

Lars Ankile et al. · arXiv 2610.05882 source

RealtimeWAM: a one-step, asynchronous-inference world action model · world-model

World action models (WAMs, which use a video-generation backbone to predict actions) are slow at inference for two reasons: actions need multi-step denoising, and the video and action experts must wait on each other in sequence. RealtimeWAM uses teacher-anchored consistency distillation to produce actions in one step, then lets the two experts share the video KV cache block by block and run in a pipelined, parallel fashion. On benchmarks such as LIBERO and RoboTwin the performance loss is under 1%, and end-to-end speed on an H100 improves by about 25x. Code and weights are released, HF 7↑.

Chengtao Lv et al. · arXiv 2610.06617 source

RACE: lengthening VLA action chunks to reduce stop-and-go · vla

Every time a VLA runs inference, the robot pauses. Longer chunks mean fewer pauses, but errors concentrate at the transitions between two sub-skills. RACE first predicts the transition moment with a single denoising step, then generates actions conditioned on that moment. On a real robot, with chunks lengthened 4x, pause time falls by about 5x, and success remains higher than fine-tuning with the same chunk length. HF 9↑.

Seonghoon Yu et al. · arXiv 2610.05719 source

What guardrails miss, robots carry out: implicit malice in VLA instructions · vla

A VLA does not refuse instructions; screening out malicious requests relies on external monitors alone. The authors hold the task fixed and vary only how explicit the malicious intent is. π0.5 executes the instruction at every level. Text guardrails can block blunt requests, but in implicit-malice episodes, in up to 95% of episodes the task is completed with no alert raised; after recalibrating for robot instructions, the figure is still as high as 90%. Linear probes can almost perfectly separate harmful from harmless instructions in the original language models and VLMs, but after robot training this separation weakens.

Sripad Karne et al. · arXiv 2610.05818 source

RobotUse: allocating compute, context and decisions for robot agents · manipulation

Existing interfaces either confine the agent to predefined tools or make it maintain lengthy execution code itself. RobotUse has the agent do only one thing, use vision to choose targets and poses, while geometry, motion planning and control are left to the backend. Sub-agents keep details inside their own subgoals, and experience is written into a persistent playbook. On RoboLab, task success is 45%, 6.7 percentage points higher than CaP-X. HF 10↑.

Junhoo Lee et al. (KAIST, Seoul National University) · arXiv 2610.04929 source

Inspect Robots: an open-source framework for evaluating embodied AI capability and safety · benchmark

General-purpose language models are increasingly able to control robot hardware directly. Inspect Robots automates the setup, execution and termination of physical evaluations, and the authors used it to evaluate six policies built on frontier language models, covering both capability and safety. Three months after release, the framework has been downloaded nearly 100,000 times.

Christopher Leet et al. · arXiv 2610.06306 source

GOTT: a contact primitive for dexterous hands that transfers across hand types · manipulation

Dexterous hands have lacked a low-level contact action that can be reused across tasks and hand types. GOTT splits manipulation into three phases, approach, establish contact and transport, and uses one feedback-based contact primitive for the middle phase on every hand type. It establishes contact reliably across many objects and several arm-hand platforms, including hand types never seen before, and its end-to-end success is higher than open-loop grasping. The authors include Pieter Abbeel.

Yulin Liu et al. · arXiv 2610.03861 source

Multiply Labs (US) | Series B | $75 million | valuation undisclosed · industrial

Led by Patrick Soon-Shiong and NantWorks. New investors include AstraZeneca, Lingotto, Teradyne and Strange Ventures, with existing backers such as Lux Capital and Founders Fund participating. The company was founded in 2016 and has raised more than $100 million in total. Multiply Labs builds biopharmaceutical manufacturing robots that automate timed, repetitive manual steps inside GMP-compliant closed environments, and its business has expanded from cell and gene therapy to antibodies, viral vectors and mRNA. The company claims it can cut per-dose cost by 74% and raise throughput up to 100x over manual work ⚠️ vendor claim. Systems are bought and operated by drugmakers themselves, not outsourced contract manufacturing. CEO Fred Parietti said: "AI-designed therapies have already multiplied past what this industry could ever manufacture."Sources: Axios source; AI Insider source

Minerva Humanoids (US) | Pre-seed | about $10 million | valuation undisclosed · humanoid

Led by General Catalyst, with Long Journey Ventures and Credo Ventures as co-leads and Hugging Face among the participants. The company was founded in late 2025 and is headquartered in San Francisco. Its first robot, Roger, is teleoperated by professionals wearing VR headsets for high-risk work such as bomb disposal and energy operations, with humans still making the key decisions. It took 5 months to go from concept to a walking prototype. The prototype was built and assembled in Germany, and the company says it wants to avoid dependence on single-source Chinese components. CEO Sandor Felber previously worked on humanoid research at MIT CSAIL and on Tesla's Optimus; CTO Maurice Rahme spent six years at Boston Dynamics and was one of the technical leads on Stretch. The first paid pilots begin this fall.Sources: AI Insider source; Dealroom source

Reactor (US) | Series A extension | $74 million raised in total | valuation undisclosed · world-model

Nvidia's venture arm NVentures and Sapphire Ventures have joined Reactor's Series A. The round was led by Lightspeed and announced in May; with these additions, total funding reaches $74 million. Reactor does not build world models itself; it sells the inference engine, GPU cloud and low-latency streaming that let world models run in real time, and claims it can reach more than 60 frames per second with latency under 40 ms ⚠️ vendor claim. Customers are currently mostly Hollywood studios, advertising and streaming. Robotics is the next focus: treating a world model as a simulation environment that responds to actions, so robot policies can be trained and evaluated in generated scenes. The new money will go to compute, team expansion, and buying robot hardware for testing. Both co-founders come from the Apple Vision Pro team, and CEO Alberto Taiuti also co-founded Luma AI and served as its CTO.Sources: Fortune source; Pulse 2.0 source

Nvidia and Foxconn publish GB300 assembly data: robot busbar installation succeeds over 95% of the time · industrial

Over 95% is the success rate of the busbar installation step when robots at Foxconn assemble test trays for Nvidia's GB300 NVL72. This step requires picking up a conductive metal bar, placing it in the tray, and tightening 16 screws, with a target time of 124 seconds. Another step, inserting four cabled connectors (two large, two small) into tight-clearance sockets, succeeds 90% to 95% of the time, with a target time of 72 seconds. The data comes from Nvidia's technical blog post "The Machines that Make the Machines". Taipei Times relayed it as test data; TechTimes says the dual-arm robots are already running on the line at Foxconn's Houston plant, using the model Isaac GR00T N1.5. Nvidia's explanation is that GB300 designs iterate quickly and run in small batches, so fixed automation does not pay off and robots that can adapt to process variation are needed. A 95% rate means roughly 1 in 20 attempts needs rework or human handover, while Nvidia's own overall yield target is 99.5%. Foxconn plans to extend such robots to its plants in Texas, Wisconsin and California.Sources: Taipei Times source; TechTimes source

Geek+ deploys more than 1,600 robots for LPP in Poland · industrial

More than 1,600 Geek+ robots are spread across two of LPP Logistics' Polish fulfillment centers, in a partnership expanded in September. Automated storage positions rise to 1.4 million, with racks 12 meters high. The RoboShuttle system and P-series picking robots handle apparel inventory and tote flow, with scheduling run by Geek+'s own robot management system. LPP Logistics handles distribution for the Gdańsk-headquartered fashion group LPP, with warehouse space of more than 700,000 square meters across three continents. LPP Logistics CEO Sebastian Sołtys said the key to the expanded partnership is that the number of robots can be flexibly scaled up or down.Source: Retail News Asia source

Einride's driverless electric trucks enter daily freight service in Ohio · autonomy

Two cabless Einride L4 electric trucks have entered daily freight service in Marysville, Ohio, running scheduled trips between several EASE Logistics warehouses, on routes that include EASE's own sites and local public roads. No one rides in the vehicles; remote operators monitor each truck off-site and take over when needed. Operations began in September, and the accumulated data on mileage, volume and uptime will be used to decide how to expand next. The project is part of the $8.8 million truck automation corridor project launched in 2021 by the Ohio Department of Transportation, DriveOhio and the Indiana Department of Transportation, and is the third autonomous truck deployment between EASE and DriveOhio. Einride CTO Henrik Green said: "Daily operations in Marysville give us the validation we need to keep expanding autonomous freight within and beyond the corridor."Source: Taiwan News (GlobeNewswire press release) source

Agility Investor Day: Digit multi-year orders exceed $300 million, realized by milestones · humanoid

Agility held an analyst and investor day in New York on October 6. Stock Titan's account that day was that multi-year orders for the Digit humanoid exceed $300 million, with the final amount depending on whether contract milestones are met. This is the figure as of May 2026 and was already disclosed when Digit 5 launched in September. TechRepublic reports that it corresponds to roughly 1,000 robots under RaaS contracts at about $8,500 per robot per month. Order value realized by milestones is not the same as recognized revenue. Digit has accumulated more than 65,000 operating hours at GXO, Schaeffler, Amazon and Toyota's Canadian plant. The company is merging with Churchill Capital Corp XI via a SPAC, expected to close in the fourth quarter, after which it will trade on Nasdaq under the ticker AGLT.Sources: Stock Titan source; Business Wire source; Agility website; TechRepublic

Exotec discloses one year of operating data from First Supply's Wisconsin facility · industrial ⚠️ vendor claim

US wholesaler First Supply's new distribution center in West Salem, Wisconsin promises that orders placed before 5 p.m. arrive the next morning, relying on Exotec's Skypod system. After the system has run for more than a year, Exotec lists these expected results: storage capacity up to 4x, picking speed 10x that of manual work, and more than 14,000 SKUs managed in the system. Next-day orders can make up as much as 20% of the center's daily order volume. First Supply's chief logistics officer Steve Baer said the system is "efficient and lets the facility expand step by step with almost no disruption".Source: Securities.io source

Saudi Arabia launches its first L4 last-mile driverless delivery in Riyadh · autonomy

Saudi Arabia's Transport General Authority has put L4 autonomous delivery vehicles into service first in Riyadh's Al Falah district to deliver orders to customers. A remote operations center monitors the vehicles in real time and can stop them immediately when necessary. The official goal is to cut "last mile" delivery costs by up to 85%; the fleet is entirely electric and supports local manufacturing of the vehicles.Source: Saudi Press Agency source

Doosan Robotics leads two Korean national physical AI projects worth 98.9 billion won in total · industrial

Government subsidies of 68.1 billion won, plus 30.8 billion won raised by Doosan itself, make up the 98.9 billion won scale of the two projects. The first project, lasting four and a half years, develops an on-device AI SoC integrating an NPU and real-time control, so that vision, force, torque and voice data are all processed on the robot itself, mainly for sensitive sectors such as defense, nuclear energy and biopharmaceuticals. Doosan will use the chip in three collaborative robots with 5, 10 and 20 kg payloads, plus an omnidirectional mobile dual-arm industrial humanoid, with commercialization targeted for 2031. The second project, lasting 45 months, uses collaborative robots, AI and digital twins to automate welding of nuclear power components. The target is a 95% match between the robot's motions and those of a skilled welder, production time cut in half, and field tests at Doosan Enerbility's nuclear equipment plant. Partner DeepX is responsible for on-device hardware and an SDK based on a Korean-made NPU, with quality validation for nuclear equipment planned for 2029 and commercialization in 2030. CEO Park In-won said labor shortages and the aging of skilled workers "cannot be solved by expanding traditional automation".Sources: The Korea Herald source; 디지털투데이 source

Unitree and AgiBot humanoids appear at South Korea's National Assembly audit; Korean officials admit physical AI investment is insufficient · humanoid ⚠️ compiled figures

One humanoid each from Unitree (Chinese humanoid maker) and AgiBot (Chinese humanoid maker) was brought on October 6 by Democratic Party of Korea lawmaker Lee Jeong-heon to the National Assembly science and technology committee's audit of the Ministry of Science and ICT, where they shook hands with lawmakers and danced as a group to music. According to data from the lawmaker's office, Unitree produced 4,200 units in 2025 and AgiBot more than 5,200; China has 160 humanoid complete-machine makers and more than 600 core component suppliers. Of South Korea's 2,814 AI companies, only 53 are hardware companies, or 1.9%, and 29 of them have never received outside investment. Deputy Prime Minister and Minister of Science and ICT Bae Gyeonghun responded: "Past investment has been far from enough." He said spending will be increased through three flagship projects. Unitree and AgiBot together exceed 9,400 units, higher than the roughly 7,000 global full-size humanoid sales for 2025 that the IFR recently published. The two sets of figures use different definitions: the former is production as given by the lawmaker's office, while the latter counts only sales of full-size models.Source: Asia Economy source

More than ten Korean care-robot companies to form a group this year and push for legal amendments · adjacent

More than 10 care-robot companies are preparing to launch a "wellness robot" council this year, with members including SK Intellix, Hyodol and Roaizen and the Korea AI Robot Industry Association as secretariat. They want to push for amending the Intelligent Robots Act: indoor care robots currently have no legal definition, and standards for safety, liability and household data are lacking. South Korea's Ministry of Health and Welfare has added a new 37.7 billion won line in next year's budget to deploy care robots in homes and institutions and support commercialization. The government also plans to introduce about 2,000 Korean-made AI robots in the public sector, 200 of them for care.Source: Seoul Economic Daily source

Critical Point (AGILINK), spun off from AgiBot, has its dexterous hand learn Su embroidery and claims more than 15,000 units delivered in total · embodied ⚠️ vendor claim

The fully direct-drive dexterous hand OmniHand 3 Ultra from Critical Point (AGILINK) completed basic steps such as twisting thread, splitting strands, threading needles and cutting thread under the guidance of Fu Xianghong, a representative inheritor of Su embroidery, adjusting force in real time using fingertip visuo-tactile sensing and reinforcement learning. This is a demonstration in a corporate promotional video, not a production-run metric. Critical Point was spun off from AgiBot's dexterous-hand business and settled in Shanghai's Xuhui district in January this year. The company says cumulative deliveries of the OmniHand series exceeded 15,000 units as of September.Sources: Sina Finance (Shanghai Observer) source; New Atlas source

San Francisco carves out no-go zones for robotaxis during Fleet Week · autonomy

On Friday through Sunday this week, from 10 a.m. to 6 p.m. local time each day, San Francisco's transportation agency will use geofencing to keep robotaxis such as Waymo out of the northern waterfront, with service halted across the entire Presidio; Uber, Lyft and taxis are not restricted. The agency cites easing congestion and protecting pedestrian safety. Previously, after the July 4 Golden Gate Bridge fireworks ended, several robotaxis stopped in traffic lanes and some ran out of battery; on the night of the citywide power outage in December 2025, city authorities at one point could not reach Waymo.Source: CBS News source

Nachi-Fujikoshi robotics revenue up 16.9% · industrial

¥21.314 billion is the robotics revenue of Japan's Nachi-Fujikoshi for December 2025 to August 2026, up 16.9% year on year and the fastest-growing of its segments; the company expects full-year robotics revenue of ¥31.4 billion, up 24.4%. Operating profit for the June-August quarter rose 94.1% year on year. The company also announced a 5-for-1 stock split.Source: BigGo Finance source