Korea to start robot foundation model R&D in 2027 with two consortia
The ministry has a structure, a first customer and a deadline it set for itself. It does not yet have a budget.
Papers, models and benchmarks that matter in embodied AI — explained for people who build and invest, not just publish.
The ministry has a structure, a first customer and a deadline it set for itself. It does not yet have a budget.
The shares are clear. The raw counts published alongside them do not reconcile, and the newest data are still filling in.
Two city-run public shows, one academic conference, two trade fairs and a sensor maker’s KOSDAQ subscription. Every date below was checked against an organiser page or Korean press as of October 2, 2026.
KAIST’s RAIBO2 quadruped finished a full marathon in 4:19:52 without recharging. The Nature paper, published September 23, reports a cost of transport of 0.25, below a human runner’s 0.37.
Skild AI says its S1 model taught itself soccer over 140+ simulated years of self-play in Isaac Sim, then played on real humanoids. Goals were the only reward in the self-play stage.
The IFR’s World Robotics 2026 report counts 5 million industrial robots in operation and 600,000+ installed in 2025. China took 59%; the US passed Japan; Korea slipped 1%.
Menlo Research released isaac_asimov, the BSD-3 Isaac Lab code that trains its $15,000 Asimov 1 humanoid to walk with PPO and adversarial motion priors. A trained checkpoint was not yet downloadable.
GR00T N1.7, Gemini Robotics 2, π0.7, Helix 2.5, GEN-1.5, Skild S1, GO-2 and UnifoLM-WLA side by side: weights, licences, parameters, robots and access, as of September 28, 2026.
No finished safety standard covers humanoids yet. ISO 25785-1 is a committee draft, China has mapped 52 standards, and makers certify against rules written for arms and wheeled carts.
Physical AI is AI that perceives, reasons and acts in the real world. A plain guide to how it differs from embodied AI and robotics, the five layers of the stack, and why Korea put a ₩3.1 trillion budget line behind the term.
At ROSCon 2026 in Toronto on September 22, Google's Intrinsic released Intrinsic Core with a CNC machine-tending reference design, and NVIDIA released Isaac ROS 5.0 with agent-ready skills.
Robocurve ran 300 hazardous-instruction trials on real robot arms. GPT-6 Astra completed 60 of its 100, Claude Fable 5.1 refused only the doll-stabbing task, and the VLA MolmoAct2 never refused.
Figure pretrained Helix 2.5 from scratch on its human-behavior dataset, Index, and reports zero-shot tidying, towel folding and bed making in 30 Bay Area homes. Without that pretraining, success fell to 9%.
The number worth keeping is not the udon. It is the move from 99% to 99.9% on a screw, which on a production line is the difference between a stop every few units and one every thousand.
The pitch is physical AI. The engineering is plumbing: Ethernet and power at the wrist, a faster controller, and fewer external PCs between a camera and the robot.
The startup from Mobile ALOHA and DexCap researchers Zipeng Fu and Chen Wang says one policy runs on arms, humanoids and mobile manipulators and learns a new task from under 30 minutes of human demonstrations.
Unitree says one checkpoint trained on about 2,500 hours of real-robot data runs 64 manipulation tasks. It was billed as open source; at launch the code, weights and data were all marked "coming soon."
Two published exploit chains give an unauthenticated attacker root on Unitree’s G1 humanoid, one of them from Bluetooth range with no pairing. The interesting part is not the intrusion. It is that the compromised endpoint has actuators.
NAVER's ARC Brain, Isaac GR00T N1.7 and Generalist's cross-embodiment model all claim the software layer that runs somebody else's robots. They disagree on what that layer is.
Grip with the left arm, hand off to the right, reorient, insert into a rack — built on Isaac GR00T N1.7 and trained by imitation learning from teleoperation. No success rate published.
J-HRTI’s Kanto Data Factory, in operation since August 26, runs collection, annotation and industrial-replica test zones on AgiBot hardware. Glove motion capture for finger-level data is planned; the consortium was founded in March 2026.
Three permits approved unanimously on August 20: Tesla up to 5,000, Waymo 1,000, Uber 1,000 via Motional and Zoox. Tesla’s Cybercab chief engineer put the realistic one-year figure at 2,500.
An automotive supplier repackages qualified driver-monitoring vision as a robot perception platform that maps people, objects and environment as one scene. No customers, pricing or timeline disclosed.
Ten times the onboard compute, 18-hour days, and an Isaac Sim plus Cosmos learning loop trained on five years of paid hospital operation across 25 hospitals.
A framework for global procurement of AutoStore systems, filed as inside information — and containing, in the company’s own words, no purchasing commitments at this time.
Wood Mackenzie models robotics as a grid load distinct from data centres. Industrial arms account for 357 TWh of the 2035 figure; humanoids account for about 6.
A3 counted 8,940 robots worth $622M in Q2. Units rose 4.3% and order value 21.3% — and the entire North American quarter is smaller than several single humanoid funding rounds this year.
MOLIT is drafting a bill to rewrite space and facility rules for delivery, parking and EV-charging robots — and 43 companies, including Samsung Electronics, Hyundai Motor and NAVER, were in the room on August 13.
VicOne turned a capture-the-flag contest into a free simulation test suite. Import your robot, run the adversarial scenarios, find out whether your vision-language model does what a manipulated camera frame tells it to.
TactaBot is three products in one: a 15-degree-of-freedom hand with no motors in it, a tactile sensor smaller than a grain of sand, and a glove that harvests skilled human hands on the factory floor. First units ship in early 2027.
Ten smart robots do not make a service. NAVER's argument is that the value sits in the layer above them — and it is now inviting other manufacturers' machines onto that layer.
Between a $30 hobby servo and a bespoke space-qualified joint there is nothing. That empty middle is a bigger constraint on robotics than most model releases.
A LexisNexis Patent Asset Index published July 29 ranks Fourier, AgiBot, LimX Dynamics, Pudu Robotics and Unitree ahead of every Western humanoid maker. The per-patent numbers tell a different story from the totals.
Google DeepMind shipped three models on July 30. The demo is a humanoid walking to a shelf. The more useful disclosure is a table showing fine manipulation succeeding 32% of the time.
The demos get the attention; the update pipeline decides whether a fleet is a business. Agency Tool Co. is betting the boring layer is the one nobody has built.
The field's flagship academic conference has stopped asking whether humanoids can work and started asking what happens to the people already doing the job — and it is putting commercial platforms on a scored task circuit alongside research robots.
Cranes, AGVs and cargo handlers that assess their surroundings and act without an operator. The interesting claim is not the automation — it is treating the whole terminal as a single embodied system.
Past the robots on the WAIC floor, the more consequential contest was in software: a new 'world model,' a shared benchmark for physical intelligence, and edge models small enough to run on a robot — China building the layer that makes hardware useful.
A new TrendForce analysis argues the humanoid contest is turning from spec sheets to supply chains — and maps who controls the parts. China is broadest, the US owns the AI brains, and Japan and Germany hold the mechanics.
The demo is a headset and a robot arm picking up a cap. The more consequential product is the rig that records what the operator's brain was doing while they demonstrated the task.
Bloomberg reports China has built 64 warehouse-scale centers — with about 20 more coming — where workers perform everyday tasks on repeat so robots can learn from them. The wager is that the humanoid race is won on demonstration data, and that China can manufacture it.
The engineering decision matters more than the demo. Physical AI went in as a module inside an existing industrial robot system, not as a replacement for it.
The numbers are self-reported and the benchmarks are the company's own. What makes them worth reading is the shape of the curve: reliability improving predictably with training time, the way language models did.
Delivery-robot maker Avride describes using cloud vision-language models not to drive, but to flag unusual scenes for human review — a pragmatic split between real-time control and cloud oversight.
Vision-language-action models turn a vision-language backbone into a robot controller that takes instructions and outputs actions. Here's why they became the field's default, in plain terms.
Training robots in simulation and transferring to hardware is getting better, cutting how much costly real-world data a policy needs. It isn't 'solved' — but the economics are shifting.
Investors are pouring money into models that predict how the physical world behaves — General Intuition's $320M round and NVIDIA's Cosmos are two faces of the same wager.
Daimon Robotics and Galbot released RobOmni, billed as the first omni-modal benchmark to include tactile sensing — a shared yardstick for the contact-rich tasks robots still fail.