Nvidia GR00T 2: 5 Demos Lift Robot Success from 40% to 95% on Unseen Tasks

On November 9, Nvidia presented GR00T 2 at a robot learning conference. The evaluation uses a dual-arm robot equipped with two Sharpa hands across 50 tasks that the robot had never seen in its own training data; those tasks appear only in human videos or other robot datasets.

Zero-shot performance

With only a language instruction and no demonstrations, the success rate is 40%. Across the 50 tasks, 20 tasks succeeded at least once. The authors note that most failures are due to a single step not being properly prepared, while the overall step sequence and order are often correct.

Few-shot adaptation with low-rank fine-tuning

Providing 1 demonstration per task raises the average to 71%. With 5 demonstrations per task (under 3 minutes of robot data per task) followed by 30 minutes of low-rank adaptation on a single RTX 5090, the average success rate reaches 95%. The same adaptation mechanism is used from 1 to 5 demonstrations — it is not a different model. After 5 demonstrations, 37 of the 50 tasks succeed on every trial, so the reported 95% is an average, not a uniform per-task figure. The research page does not state how many trials were run per task.

Pre-training data

The base model is pre-trained on three data sources: 402,730 hours of human first-person (egocentric) video, 10,184 hours of hand-held gripper demonstrations, and 23,500 hours of robot teleoperation data. The research page notes that 100,000 hours of the human video are available on Hugging Face. Model weights for GR00T 2 have not been released.

Architecture and learning mechanism

The model consumes the most recent 0.8 seconds from three cameras, a language instruction, and the current joint state. It jointly denoises future video frames and a sequence of joint targets, then executes and observes the new frames. Predicting future frames is what enables the model to learn "how the scene changes when the hand moves" from human video data. Few-shot adaptation modifies only the low-rank adaptation branches; the three pre-training datasets are not re-trained.

Additional notes

The research page also shows memory insertion, toy assembly, and chaining with another model, but those success rates come from different tasks and trial counts and cannot be compared directly to the 50-task 95% figure. A smaller Turbo variant is listed at 2 billion parameters; the parameter count for the main model is not disclosed.

Research page: https://research.nvidia.com/labs/gear/gr00t-2/

Code example

👀点击关注
AI前沿快速Follow