| smoke_config | groot.yaml |
|---|
Dual-system VLA from NVIDIA. Paper | GitHub
| File | Benchmark | Checkpoint |
|---|---|---|
groot.yaml |
Generic | nvidia/GR00T-N1.6-3B |
libero.yaml |
LIBERO | nvidia/GR00T-N1.6-3B |
simpler_widowx.yaml |
SimplerEnv WidowX | nvidia/GR00T-N1.6-bridge |
simpler_google_robot.yaml |
SimplerEnv GR | nvidia/GR00T-N1.6-fractal |
robocasa365.yaml |
RoboCasa365 | robocasa/robocasa365_checkpoints (N1.5) |
robocasa365.yaml runs the benchmark's own N1.5 baseline, so it uses a
separate script pinned to the RoboCasa fork of Isaac-GR00T rather than the
groot.py server. Set ROBOCASA365_GROOT_CKPT to the
gr00t_n1-5/multitask_learning/checkpoint-120000 directory. That fork's
bundled Eagle backbone imports FlashAttention directly, so the script pins the
exercised FlashAttention build instead of using the Transformers kernels
extra.