How robot foundation models learn from one demonstration

Teaching a robot a new job has usually meant weeks of programming or thousands of recorded examples. This week NVIDIA claimed a large step away from that. At the Conference on Robot Learning (CoRL) in Austin, Texas, it announced GR00T 2, a robot model that, in NVIDIA's own tests, reached a 71% average success rate on 50 unfamiliar tasks after one demonstration of each, and 95% after five. Alongside it came Isaac Lab 3.0, a simulation framework that can switch between two physics engines.

Robots are also leaving the lab in a more ordinary way. ANYbotics began delivering ANYmal X, a legged inspection robot built to the ATEX and IECEx explosion-protection standards, to Equinor and DSM-Firmenich. Boston Dynamics put scooter wheels on Atlas to test how well controllers trained in simulation work on real hardware.

These stories sit at two ends of one pipeline. At one end, foundation models learn general skills from video and robot data. At the other, a robot must pass certification and deliver useful data every day in a plant. This article covers both.

What happened

NVIDIA announces GR00T 2 and Isaac Lab 3.0 at CoRL (9 November 2026)

CoRL 2026 runs from 9 to 12 November in Austin. On the opening day NVIDIA introduced GR00T 2, which it calls an open "world action model" for humanoid robots. The Next Web reported that NVIDIA is also releasing 100,000 hours of first-person (egocentric) human video for anyone to use.

According to NVIDIA's whitepaper, GR00T 2 is a single transformer of about 2 billion parameters, built on the Cosmos-Predict2.5-2B video model. It predicts future camera frames and robot actions together. A faster variant, GR00T 2 Turbo, makes a prediction in 124 milliseconds on NVIDIA's Jetson AGX Thor embedded computer.

The headline numbers come from a test called Dexter-50: 50 tasks on a two-armed robot with two 22-joint hands, none of which appear in that robot's training data. With only a language instruction, GR00T 2 succeeded in 40% of trials and completed 40 of the 50 tasks at least once. With one demonstration per task it averaged 71%. With five, about three minutes of robot data, it averaged 95%. Each adaptation takes about 30 minutes on one RTX 5090 graphics card, the paper says.

NVIDIA said Foxconn, Siemens, Nexuni and MicroAGI are testing or using the model, for uses including server manufacturing, laundry handling and, at MicroAGI, Ethernet cable swapping on a YAM robot arm. The Next Web noted that none of these companies has said it will use GR00T 2 in production. Reports gave no release date or licence for the weights.

Isaac Lab 3.0 is NVIDIA's open framework for training robots in simulation. According to NVIDIA's announcement, as summarised by HyperAI, it can switch between the PhysX and Newton physics solvers, lets AI agents configure training environments, and supports reinforcement learning post-training of models such as GR00T. NVIDIA says deeper Newton integration improves simulation of cables, cloth and robot hands, and reports "up to 2.5 times the training throughput" of Isaac Lab 2.3 with PhysX on one dexterous manipulation benchmark.

NVIDIA also presented research. GRAIL, from NVIDIA and UCLA and a CoRL Best Paper Award finalist, generates humanoid training data from 3D object assets and a video world model, without teleoperation or motion capture. RoboTTT, from NVIDIA, Stanford and UT Austin, lets a robot policy use about five minutes of history (8,000 steps at 30 Hz), and its authors say it does so without increasing inference latency.

ANYbotics begins deliveries of ANYmal X (27 October 2026)

On 27 October 2026 Swiss company ANYbotics said it had begun delivering ANYmal X to oil, gas and chemical companies, including Equinor and DSM-Firmenich. SiliconANGLE reported that ANYbotics calls it the first legged robot designed for areas where a spark could set off an explosion.

According to World Pipelines, its computer and inspection equipment sit in sealed, nitrogen-filled chambers, its twelve leg actuators and battery have flameproof casings, and conductive feet ground it. It carries thermal, visual, acoustic and gas sensors, climbs steep stairs and docks itself to charge. Its Shift software turns findings into maintenance work orders and connects to SAP, Yokogawa and SLB plant systems. BP, Shell, BASF and others have tested it since 2023. ANYbotics says more than 50 customers run over 120,000 autonomous inspections a month on its wider system; it did not say how many use ANYmal X.

Boston Dynamics puts Atlas on wheels (9 November 2026)

Boston Dynamics published a video of its research Atlas humanoid with its feet replaced by standard electric scooter wheels, fitted to the ankle actuators with a custom adapter, as reported by Interesting Engineering. Control was trained with reinforcement learning in simulation after engineers matched the simulated motor models to measured actuator behaviour. A 360-degree spinning jump, following recorded human motion, worked on the first attempt. The company says the wheeled Atlas is a research platform, not a planned product.

From language models to robot models

A vision-language-action model (VLA) is the robot version of a chatbot. A chatbot takes text and returns text. A VLA takes camera images and an instruction, such as "put the cup in the sink", and returns actions, such as joint positions and gripper commands, many times a second. It usually starts from a vision-language model that already knows what cups and sinks look like, then learns to turn that knowledge into motion.

Think of a new kitchen assistant. They already know what a cup is. What they still need to learn is how hard to grip this cup and where this sink is. The vision-language part supplies the general knowledge; the action part supplies the hands.

What a world action model adds

NVIDIA calls GR00T 2 a world action model because it predicts the next actions and what the cameras will see next. A world model can imagine how a scene will change. Learning to predict future video forces a model to pick up some physics: objects fall, cloth folds, a drawer stops at its end.

It is the difference between someone who has memorised a recipe's steps and someone who can picture what the pan will look like in 30 seconds. The second notices sooner when something goes wrong. Because GR00T 2 starts from a video generator, it inherits a lot of knowledge about how the visual world moves.

The paper is also candid about limits. For recovery from messy starting positions, GR00T 2 relies on a separate agent that checks and repairs progress. That agent's median turn takes 7.5 seconds, and in a median episode the robot waits for it 38% of the time.

How robot foundation models are trained

Language models learned from the internet's text. There is no internet of robot motion, so builders mix several kinds of data. The GR00T 2 whitepaper lists its sources.

Data source	What it is	GR00T 2 amount	Strength	Weakness
Human egocentric video	People filmed from a head camera	402,730 hours	Huge variety, cheap	No robot joint data
Handheld gripper (UMI)	People use a hand-held gripper with a camera	10,184 hours	Close to robot grasps	Still no robot body
Teleoperation	A person remote-controls a real robot	About 23,500 hours	Exact robot actions	Slow and expensive
Simulation	Virtual robots in physics engines	Used for fine-tuning	Unlimited and safe	The reality gap

By these totals, human data makes up roughly nine-tenths of the training hours. NVIDIA is betting that a model can learn what tasks look like from people and needs only a little robot data to learn how its own body should do them.

Teleoperation is like teaching a child to write by guiding their hand: exact, but slow. An operator drives the robot with a headset, gloves or a matching "leader" arm, and every movement is recorded. Each hour costs an hour of skilled time, which is why the largest sets are tens of thousands of hours, not millions.

Human video is plentiful but indirect. A model sees hands open a jar but not the robot's motor commands, so it must map "what success looks like" onto its own body. It is closer to learning a dance from a video than from a teacher holding your arms.

Simulation and sim-to-real

In simulation a robot can practise a million times overnight without breaking anything. The difficulty is the reality gap: friction, lighting, cable stiffness and motor delays are never quite right. Two techniques help. Domain randomisation varies colours, lighting, masses and friction so the real world looks like one more variation. System identification measures the real hardware and tunes the simulation to match, which is what Boston Dynamics did before training the wheeled Atlas.

Simulation is also where reinforcement learning works best: the robot tries, gets a reward and improves. Isaac Lab 3.0's support for post-training foundation models this way borrows a pattern from language models: pretrain by imitation, then refine by trial and error.

GRAIL shows another route, synthetic data. It uses a video world model to imagine a person handling a known 3D object, recovers the 3D motion and converts it into training data for a Unitree G1 humanoid. Its authors report that in early GR00T co-training, a mix of 95% GRAIL data and 5% teleoperation improved grasping success over teleoperation alone.

One-shot imitation and how to read the numbers

Imitation learning means learning by copying demonstrations. One-shot means adapting to a new task from one demonstration, few-shot from a handful, and zero-shot from an instruction alone.

It works for the same reason it works in language models: a pretrained model already knows a lot, so one example points it at the right behaviour. An experienced cook who watches you make your family's dumplings once can usually make them. Someone who has never cooked cannot. GR00T 2 uses the demonstration to fine-tune itself with LoRA, which trains a small set of extra weights instead of the whole network. That is why adaptation takes half an hour on one card rather than days on a cluster.

Reading success rates carefully

Setting	Robot data per task	Reported Dexter-50 result
Zero-shot	None, instruction only	40% of trials; 40 of 50 tasks done at least once
One-shot	1 demonstration	71% average success
Five-shot	5 demonstrations (about 3 minutes)	95% average; 37 of 50 tasks succeeded every trial

A success rate looks simple but hides choices. Ask four questions of any robot benchmark.

How were the tasks chosen? The Dexter-50 tasks are absent from that robot's data but present in other robots' data. A figure caption says the tasks are "those the model learns with five episodes (at least 80% success)". So 95% is not the rate for any task you might bring.

Is it capability or reliability? Completing a task "at least once" shows capability. A factory needs the same task to succeed thousands of times in a row.

How many trials, judged by whom? Small trial counts mean wide error bars. Look for blind judging and published trial counts.

Compared with what? On six long tasks with a YAM arm, the paper reports 71.7% for GR00T 2 against 12.5% for π0.5 and 8.3% for GR00T 1.7. These are the authors' own runs of competitors.

As with any model score, a number is only as useful as the description of how it was measured; see a score is only as good as its report.

Why simulators and physics solvers matter

A physics solver is the part of a simulator that works out how things move and collide at each tiny time step. It is the referee of practice: if its rules are wrong, what the robot learns does not carry over to the real match.

Solvers make different trade-offs. PhysX is NVIDIA's long-standing engine, widely used for rigid bodies. Newton is a newer open-source engine aimed at robotics, and NVIDIA says it handles deformable objects and articulated hands better in Isaac Lab 3.0. That matters because robots now attempt floppy things: MicroAGI's Ethernet cable and Nexuni's laundry are hard to simulate with an engine built mainly for rigid boxes.

Switching solvers in one framework is useful for a second reason. If a skill works only under one engine, it may be exploiting that engine's quirks, like a student who has learned one teacher's exams rather than the subject. Testing under both is a cheap check.

Speed matters too, because reinforcement learning needs millions of attempts. NVIDIA's AlphaNIST example trained one policy in Isaac Lab for all 20 tasks of the NIST Assembly Task Board, reporting over 98% training success against 81% for per-task policies, with about 39 times less GPU time. That is success in simulation, not on a real line. Simulation also serves as a test bench, close to the idea of a digital twin: a virtual copy of a cell where changes are tried before they reach the floor.

What it means in practice

Adaptation cost is falling. If NVIDIA's results hold up elsewhere, a task that once needed hundreds of demonstrations might need minutes of recording and half an hour of training. That suits short-run, varied jobs that rigid automation handles badly.

Industrial robots ship on certification, not demos. ANYmal X sells on enclosures, flameproof actuators and SAP integration, not success percentages. As ANYbotics' chief revenue officer put it: "A certified robot is only the entry ticket for inspecting oil, gas, and chemicals facilities."

Inspection comes first. Walking a route, reading gauges, listening for abnormal noise and detecting gas are repetitive, well defined and dangerous for people. They need reliable navigation more than dexterous hands, which is why legged inspection robots are deployed today while dexterous humanoids are still in testing. Their data feeds the predictive maintenance systems many plants already run.

Deployment means safety engineering. A robot in a hazardous area must not ignite the atmosphere; a robot in a factory must not hurt people. That takes risk assessments, certified components, defined zones, emergency stops and a plan for failure. A policy with 95% success still fails one attempt in twenty, so the cell around it must make failure safe and recoverable. See also smart manufacturing and Industry 4.0.

What to do now

Start with inspection. If you run a plant, look at inspection rounds, where certified products exist, before dexterous robot work.
Sort tasks by variety and volume. Few-shot models suit many varied, short-run tasks; fixed automation still wins on one task repeated millions of times.
Record good demonstrations now. Clean, labelled video or teleoperation of your own tasks will be the "one shot" for future models. Note camera positions and lighting.
Build an evaluation cell. Fix tasks, trial counts, success criteria and a blind judge before testing any vendor model.
Ask vendors how tasks were chosen. Request results on your tasks with trial counts, failure modes and human interventions.
Model the cell in simulation. A simple digital twin lets you test policies, layouts and safety zones before touching hardware.
Check certification early. Confirm ATEX or IECEx coverage for the exact zone, and involve your safety engineer before the pilot.
Plan the data path. Decide where findings go, such as the maintenance system, so inspections lead to work orders.
Design for failure. Specify what the robot does when unsure or stuck, and who is called.

Risks and open questions

Results are the vendor's own. GR00T 2's numbers come from NVIDIA's paper and robots. Independent replication on other hardware has not been published.

Task selection flatters averages. The Dexter-50 tasks were filtered to ones the model learns well. Results on arbitrary tasks will be lower.

Availability is unclear. NVIDIA calls GR00T 2 open, but reports did not say when or under what licence the weights will be downloadable.

Touch and checking are weak. The paper lists force and touch as open problems. It also reports that in three of eight failures on one task, both the model and its agent judged a screw fully driven when it was not.

Human video raises consent questions. Egocentric datasets show homes and workplaces. Ask how footage was collected, consented and filtered.

Certification lags learning. Safety standards assume predictable behaviour. How to certify a policy that changes after each fine-tune is not settled.

Watch next: independent tests of GR00T 2 on other robots, its licence and release date, whether Foxconn or Siemens move from testing to production, ANYmal X's public showing at ADIPEC 2026, and the CoRL best paper decision.

Frequently asked questions

What is a vision-language-action model?

It takes camera images and an instruction and outputs robot actions. It usually starts from a vision-language model that understands objects and words, then learns to map that to motor commands. GR00T 2 also predicts future video.

Can GR00T 2 really learn a task from one demonstration?

In NVIDIA's Dexter-50 test, one demonstration plus about 30 minutes of LoRA fine-tuning gave 71% average success. The tasks appeared in other robots' data and were selected from tasks the model learns well. Expect lower results on unrelated tasks.

Why train robots on human video?

It is far cheaper and more varied than robot data. In GR00T 2's training, human data made up roughly nine-tenths of the hours. The model learns what tasks look like from people and needs little robot data to learn its own movements.

What does ATEX and IECEx certification mean for a robot?

They are explosion-protection standards for equipment used where flammable gas or dust may be present. The robot must not create sparks or hot surfaces that could ignite the atmosphere. ANYmal X uses sealed nitrogen-filled enclosures, flameproof actuators and grounding.

Is the wheeled Atlas going on sale?

No. Boston Dynamics calls it an experimental research platform for testing how well simulation-trained controllers run on real hardware. It says the findings will inform the standard Atlas.

Summary

At CoRL 2026 NVIDIA announced GR00T 2, a 2-billion-parameter world action model trained mostly on human video, which in NVIDIA's tests reached 71% average success on 50 new tasks after one demonstration and 95% after five. Isaac Lab 3.0 adds switchable PhysX and Newton physics and reinforcement learning post-training, while GRAIL and RoboTTT push synthetic data and long memory. ANYbotics began shipping the explosion-protected ANYmal X to Equinor and DSM-Firmenich, and Boston Dynamics tested sim-to-real control on a wheeled Atlas. Robot models are learning faster, but their success rates need careful reading, and real deployment still depends on certification, safety design and integration.

GR00T 2 and Isaac Lab 3.0: robots that learn a new task from one demonstration