By Paul Derrington, Jr.
CEO | Chief Engineer | Research Engineer | Editor, D-CSIL
4 min read
Key Takeaways
NVIDIA unveiled GR00T 2 on Monday at the Conference on Robot Learning in Austin — a robot foundation model that learns new manipulation skills mostly by watching video of people doing the job. On 50 tasks the target robot had never practiced, it succeeded 40% of the time with no demonstrations at all, 71% with one demonstration, and 95% with five. NVIDIA also released 100,000 hours of the human video it trained on, and a faster variant, GR00T 2 Turbo, that runs onboard a Unitree G1 humanoid.
Watch first, move second: what a world action model is
GR00T 2 is a world action model: one diffusion transformer that predicts the next few frames of video and the robot's next actions together. Instead of mapping camera input straight to motor commands, it imagines how the scene should change as it acts — which is how it learns physics from watching people. It reads recent video, the instruction, and the robot's state, denoises a future video and an action chunk together, executes the chunk, and repeats. The same model runs on parallel grippers, dexterous hands, and humanoids, with small per-robot heads turning the plan into motor commands for each body.
GR00T 2 is pretrained on 402,730 hours of egocentric video of people at work, 10,184 hours of demonstrations with a handheld UMI gripper across 2,897 tasks, and 23,500 hours of robot teleoperation. Human video is by far the largest share, and NVIDIA measured the payoff directly: pretraining on all of the human video instead of none raised real-robot task progress from 50% to 70% across 72 Dexter tasks.
The worked example: five Ethernet cables, one robot arm
NVIDIA's research page describes an agent setup: a reasoning agent breaks the work into steps and watches progress; GR00T 2 does the contact-rich insertion itself; when a cable drops or a step stalls, the agent repairs the scene with a coded skill and hands control back. In NVIDIA's tests, with the RL-trained GR00T 2, that setup inserted five Ethernet cables into a network switch in 19 of 20 runs, and plugged five chargers into pointed-at outlets in 20 of 20 — against 12 of 20 and 2 of 20 with the supervised policy in its place. The same pattern held on toy assembly: the YAM arm put together a toy plane, a toy tractor, and a Hot Wheels car in 98, 97, and 97 of 100 trials, each trial starting with the parts dumped in a cluttered pile.
Shown a task only by gloved people wearing head and wrist cameras — no robot teleoperation at all — a two-handed robot lined a lab funnel with filter paper and wet it with a wash bottle in 65% of trials. Shown only with a handheld gripper, the YAM arm assembled a toy dump truck end to end in 60% of trials, a job that takes up to four minutes.
The numbers: 40 of 50 tasks with no demos, 95% with five
The headline evaluation is 50 tasks the target robot — a bimanual machine with two 22-degree-of-freedom Sharpa hands — had never practiced, seen only in human video or on other robots. Zero-shot, GR00T 2 succeeded 40% of the time and completed 40 of the 50 tasks at least once. One demonstration per task lifted average success to 71%; five demonstrations — under three minutes of robot data — lifted it to 95%, with 37 of the 50 tasks succeeding in every trial. Each adaptation is a 30-minute LoRA fine-tune on a single RTX 5090.
With minimal post-training it seats GB300 memory modules with two 22-DoF hands 85% of the time and assembles a toy truck 65% of the time — tasks NVIDIA says previous models could not do. It also ranks among the top three models in every vision-language-action and world action model category on the RoboCasa365, RoboLab, RoboDojo, and MolmoSpaces benchmarks, with multiple leaderboard wins. All of these are NVIDIA's own figures from its research page and blog; no independent replication has been published yet.
What you can use today — and the open question
The 100,000 hours of human egocentric video is out now as Human-1M-100K on Hugging Face, openly available under a CC BY 4.0 license. The model itself is a different story: NVIDIA calls GR00T 2 open, but the weights and code are slated to follow in December, according to coverage of the CoRL presentation — so for now the release is the data, the research page, and the paper, not a model you can download. Isaac Lab 3.0, NVIDIA's open-source robot-learning framework, is available now; NVIDIA says that with the Newton physics engine it reached up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX on a vision-based dexterous manipulation benchmark. Two new open benchmarks, DexBench (15 tasks, four scenes) and VABAR (six industrial tasks, from UC Berkeley and NVIDIA), are also new.
Early industrial interest is testing, not deployment: Nexuni is trying it for robotic laundry handling, Foxconn is applying it to server-manufacturing automation, and Siemens is evaluating it across industrial uses. UC Berkeley robotics professor Ken Goldberg, a lead on the VABAR and Graph-as-Policy projects, put it this way: "To me, agentic robotics is the most exciting technological development since the internet in 1995."