GR00T-N1.6 inference for MESA / BiMESA

This is a minimal, inference-only fork of NVIDIA Isaac-GR00T (N1.6, upstream commit d331b68) for serving the GR00T-N1.6 baselines from the MESA paper to the MESA evaluation server. Training code, simulation harnesses, and examples from the upstream repository have been removed; see upstream for training and finetuning.

Changes relative to upstream:

• gr00t/eval/serve_mesa.py: websocket policy server speaking the openpi-client protocol used by MESA's scripts/eval_server_parallel.py. Camera names, state layout, and action keys are read from the checkpoint.
• gr00t/data/state_action/state_action_processor.py: MESA/BiMESA relative joint actions keep the gripper dimensions absolute (_mesa_joint_reference). The released checkpoints were trained with this convention.
• pyproject.toml / uv.lock: transformers==4.51.3 (the version used to train and evaluate the checkpoints) and websockets.

Installation

Requires Python 3.10, CUDA 12, and uv.

git clone https://github.com/pairlab/mesa-GR00T.git && cd mesa-GR00T
uv sync

Serving a checkpoint

MESA and BiMESA GR00T-N1.6 checkpoints are finetuned from nvidia/GR00T-N1.6-3B with the new_embodiment tag and store their modality configuration and normalization statistics in the checkpoint directory:

Checkpoint	Setting
albertwilcox/mesa-gr00t-n1.6	single-arm MESA (trained on MESA-70)
albertwilcox/bimesa-gr00t-n1.6	bimanual BiMESA (trained on BiMESA-57)

Download a checkpoint and start the policy server:

uv run huggingface-cli download albertwilcox/bimesa-gr00t-n1.6 --local-dir checkpoints/bimesa-gr00t-n1.6
uv run python gr00t/eval/serve_mesa.py --model-path checkpoints/bimesa-gr00t-n1.6 --port 8001

Then, from the MESA repository, run the evaluation server against the same port.

Single-arm MESA (Franka; left-shoulder + wrist cameras; 8-D joint-position state/action):

uv run scripts/eval_server_parallel.py --port 8001 --eval-set-name mesa-70 \
  --num-rollouts-per-task 50 --controller-type joint_pos

Bimanual BiMESA (two ReverseMountedYam arms; egocentric + two wrist cameras; 14-D joint-position state/action):

uv run scripts/eval_server_parallel.py --port 8001 --eval-set-name bimesa-id \
  --num-rollouts-per-task 50 --controller-type joint_pos \
  --robots ReverseMountedYam ReverseMountedYam \
  --camera-names egocentric robot0_eye_in_hand robot1_eye_in_hand \
  --state-keys robot0_joint_pos robot0_gripper_jaw_width robot1_joint_pos robot1_gripper_jaw_width

The policy predicts 20-step action chunks at 20 Hz; the MESA evaluation server executes the first 5 actions before replanning.

License

Code is released under the upstream Isaac-GR00T license. GR00T-N1.6 model weights are subject to NVIDIA's model license; see nvidia/GR00T-N1.6-3B.