rohansiva/gr00t-libero-x

GR00T N1.7 fine-tuned on LIBERO-X (Levels 1-3, 60-task subset)

Fine-tuned from nvidia/GR00T-N1.7-LIBERO (libero_10 checkpoint) on a subset of the LIBERO-X dataset.

Training data

• 60 tasks sampled (seed=42, no cross-level overlap) from LIBERO-X's Level 1-3 evaluation-protocol task pools (20 per level), each requiring at least 1 recorded demo episode
• 372 episodes / 123,224 frames total
• Embodiment: LIBERO_PANDA

Training setup

• Base: nvidia/GR00T-N1.7-LIBERO (libero_10)
• Trainable: projector + diffusion action head only (1.62B / 3.14B params, 51.5%) — VLM backbone (nvidia/Cosmos-Reason2-2B) frozen
• 8x GPU, DeepSpeed ZeRO-2
• Global batch size: 192
• 15 epochs (9,627 steps)
• Learning rate: 3e-5 (cosine, linear-scaled from the published recipe's 1e-4 @ batch 640)
• Warmup ratio: 0.05, weight decay: 1e-5, state dropout: 0.2

Final train loss: 0.177 (from ~1.35 at start).

Evaluation

Open-loop action-prediction MSE/MAE against 11 held-out trajectories spanning 30 tasks (Levels 1-3, disjoint from the 60 training tasks):

• Average MSE: 0.0357
• Average MAE: 0.0894

This measures action-prediction accuracy against recorded ground-truth trajectories, not closed-loop task success rate in simulation.