franka-gr00t-scripts

franka-gr00t-scripts

GR00T-N1.5 real-robot deployment, reference inference, and data post-processing scripts for the Franka end-effector (DROID) rig.

Layout

my_scripts/ — Franka-EEF deploy (this rig, tactile-free)

• deploy_franka_eef.py — main real-robot driver. cartesian_position control, absolute EEF targets. Adds --task (non-interactive instruction), --max_trans_per_step (per-step move clamp vs. measured EEF), home-at-start/finish, and opens only the two cameras named by --left_camera_id/--wrist_camera_id.
• deploy_franka_eef_velocity.py — variant using cartesian_velocity + clip(action,-1,1) (the "standard" GR00T/openpi path). Note: this checkpoint outputs absolute positions, so velocity mode drifts — for observation only.
• full_inference_mock.py — full closed-loop rollout with the robot/cameras mocked (verify the pipeline without hardware).
• run_example.sh, example.txt, README.md — usage + rig notes.

john_scripts/

• groot_policy_inference.py — John's groot_dust deploy driver (joint-position control, measured-position safety clamp calibrated from real closed-loop data).

default_inference/

• gr00t_action_quantization/ — the base DROID driver my_scripts was forked from (groot_policy_inference.py + launch .shs).
• nvidia_mygr00t/ — upstream NVIDIA reference: inference_service.py, serve_policy.py, and deployment_scripts/ (ONNX/TensorRT export + inference).

lerobot_conversion/ — data post-processing

• convert_teleop_to_lerobot__jonghoon.py / ..._cartesian_position__jonghoon.py — teleop → LeRobot converters for the Franka rig.
• convert_teleop_to_lerobot_pi05__jimin.py — jimin's pi0.5 converter.
• extract_video.py, concat_lerobot_files.py — video extraction / shard concatenation.

Notes

• Model weights / FLARE-token differences are out of scope here — this repo is code only.
• Authors: deploy/data scripts by the rlwrld lab (jimin, jonghoon, john); NVIDIA reference under default_inference/nvidia_mygr00t/ retains its Apache-2.0 headers.