PhAIL – NVIDIA GR00T

Model card

Made by: NVIDIA
Size: 3B parameters
Published: 18 Mar 2025 (GR00T N1)

More about the model

Publication: arXiv 2503.14734 (GR00T N1)
Code: github.com/NVIDIA/Isaac-GR00T
Weights licence: N1.6: NVIDIA License, research and evaluation only source; N1.7: NVIDIA Open Model License source
Actions per chunk: 16 in N1.6, and 40 in N1.7 source
Smallest GPU: Jetson Orin, the smallest device in NVIDIA's timing table source
Inference time: N1.6: 44 ms on an RTX 4090 (torch.compile, 4 denoising steps, one camera) source; N1.7: 48.6 ms on an H100 (torch.compile, 4 denoising steps, one camera) source

Architecture: Vision-language model and a flow-matching diffusion transformer. N1.6 uses Eagle-Block2A-2B-v2, and N1.7 uses Cosmos-Reason2-2B (Qwen3-VL) source
Training data: N1.6: base model, pre-trained on 10k+ hours of robot data source; N1.7 DROID: post-trained on DROID, about 76,000 trajectories (350 hours) source

Run with Positronic

GR00T N1.7 DROID

Start the server source

docker run --gpus all -p 8000:8000 positro/gr00t-n17-droid

Drive a DROID robot from a positronic checkout guide

uv run --extra hardware positronic-inference real \
  --policy=.remote \
  --policy.address.host=localhost --policy.address.port=8000 \
  --next_task.instruction="Put the marker on the plate"

Progress through an episode

Each band shows how many episodes of GR00T were in that band at each second. On a task of one step, an episode is in the band of the furthest stage it reached. On a task of several steps, an episode is in the band its credit has reached, and each band starts at its stage's credit for one step. The dashed line shows the mean Reliability.

Episodes

Each episode of GR00T on DROID: Everyday tasks, the newest first.

Episode details

Progress through a run

Each band shows how many runs of GR00T have placed that number of items at each second. The dashed line shows the mean share of placed items per run.

Loading runs...

Episodes

Each episode of GR00T on Pick and Place, the newest first.

Episode details

Progress through an episode

Each band shows how many episodes of GR00T had succeeded, partly succeeded (some subtask score) or had no success at each second of simulated time. An episode that has ended stays in its band. The dashed line shows the mean Reliability, the mean subtask score.

Episodes

Each episode of GR00T on RoboLab-120, the newest first.

Episode details

Progress through an episode

Each band shows how many episodes of GR00T had succeeded or had no success at each second of simulated time. An episode that has ended stays in its band. The dashed line shows the mean Reliability, the share of the episodes that had succeeded.

Episodes

Each episode of GR00T on MolmoSpaces-320, the newest first.

Episode details