PID0930/gr00t-n17-pick-up-corn

GR00T N1.7 - pick_up_corn (Franka FR3 + Inspire RH56)

Fine-tuned from nvidia/GR00T-N1.7-3B on the pick_up_corn dataset (30 episodes, ~17k frames). Inference-only weights: DeepSpeed optimizer/scheduler/RNG states are intentionally excluded.

• Embodiment tag: NEW_EMBODIMENT (new_embodiment)
• Task / prompt: "Pick up the corn and place it in the cup"
• Robot: Franka FR3 (7-DOF arm) + Inspire RH56 (6-DOF hand)
• Cameras: tpv, third_person, wrist (480x640)
• Training: ~24,000 steps, global batch 64, 8x A100-80GB

Modality

• state: arm(7) + hand(6)
• action: arm(7, RELATIVE) + hand(6, ABSOLUTE), 16-step prediction horizon
• video: tpv, third_person, wrist

Load (Isaac-GR00T)

python gr00t/eval/run_gr00t_server.py --model-path <this repo> --embodiment-tag new_embodiment

Use the matching modality config (state=arm+hand, action=arm+hand, video=tpv/third_person/wrist).