GR00T N1.7 TAPTIC 11 tasks
Collection
GR00T N1.7 fine-tuned on 11 TAPTIC Panda+XHand tasks: dataset (private) and best checkpoints. • 2 items • Updated
GR00T N1.7, TAPTIC 11 tasks (Panda + XHand)
Fine-tuning of nvidia/GR00T-N1.7-3B with lerobot 0.6.2 on 11 TAPTIC tasks (private dataset
TAPTIC/multitask_11tasks_reduced: 715 episodes, 239k frames, 10 fps; 10% of the episodes held out for evaluation).
- input: 2 cameras (wrist, left) + 40-dim state (7 Panda joints, 12 XHand joints, 15 finger forces, ee pose 6); output: 18-dim action (ee pose 6 + 12 XHand joints), chunk 40
- batch 128, 20k steps (~12 epochs), lr 1e-4 cosine (planned for 10k, stretched to 20k when resumed at step 6k), VLM frozen, image augmentation
- episode tails included in training (lerobot drop_n_last_frames set to 0)
- uploaded checkpoints, open-loop error on the held-out episodes (first 16 actions of each chunk):
step-020000: 31.94 mm, 6.03 deg, hand 5.32 degstep-018000: 32.05 mm, 6.07 deg, hand 5.36 degstep-006000: 34.97 mm, 6.33 deg, hand 5.33 deg
main=step-020000
from lerobot.policies.groot.modeling_groot import GrootPolicy
policy = GrootPolicy.from_pretrained("TAPTIC/groot17_11tasks_b128_20k", revision="step-020000")
- Downloads last month
- -
Model tree for TAPTIC/groot17_11tasks_b128_20k
Base model
nvidia/GR00T-N1.7-3B