quangnd58/GR00T-N1.7-3B-so101-multitask

GR00T-N1.7-3B finetuned on SO-101 multi-task

Finetune of nvidia/GR00T-N1.7-3B on a 60-episode SO-101 teleoperation dataset covering 3 pick-and-place tasks: "Put the cup into the box", "Put the tape into the box", "Put the tape into the cup".

Same dataset, same modality config, and same training recipe as quangnd58/GR00T-N1.6-3B-so101-multitask — this is the N1.7 counterpart, for comparison against the N1.6 backbone.

• —Code: NVIDIA/Isaac-GR00T, main branch (N1.7)
• —Embodiment tag: NEW_EMBODIMENT
• —Training: 5000 steps, global batch size 32, single H100, gr00t/experiment/launch_finetune.py
• —Final train loss: 0.0398 (train_loss avg 0.0946)
• —Open-loop eval (on training trajectories, no held-out split): MSE 8.8-24.5, MAE 1.84-2.80 deg across 6 trajectories (avg MSE 15.1, avg MAE 2.18)

Modality layout (differs from the stock SO100/SO101 example)

wrist_roll was split out of single_arm into its own ABSOLUTE action block, instead of staying inside the RELATIVE single_arm block as in NVIDIA's examples/SO100/so100_config.py. Reason: in this dataset the follower's wrist_roll state does not track the commanded action (corr = -0.038 across all 21,104 frames; every other joint correlates at 0.906-0.994). Keeping it RELATIVE would train the target as action - state against an unrelated reading. This mechanism is identical in N1.6 and N1.7 — only the base model backbone differs. See experiment_cfg/final_processor_config.json for the exact modality config baked into this checkpoint.

License

Weights inherit NVIDIA's One-Way Noncommercial License from the base model. Non-commercial use only.