gr00t-libero-x-90task

GR00T N1.7 fine-tuned on LIBERO-X (Levels 1-3, 90-task subset)

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

Successor to rohansiva/gr00t-libero-x (60 tasks, 15 epochs) — wider task coverage and more epochs, trained per team discussion to prioritize success rate on a fixed task set over held-out generalization (that generalization work moves to a downstream probe stage instead).

Training data

• 90 tasks sampled (seed=42, no cross-level overlap) from LIBERO-X's Level 1-3 evaluation-protocol task pools (30 per level), each requiring at least 1 recorded demo episode
• 625 episodes / 208,964 frames total
• Embodiment: LIBERO_PANDA
• Train == eval task set (intentional — this run targets decent success rate on these specific 90 tasks, not held-out generalization; that's tested separately at the downstream probe stage)

Training setup

• Base: nvidia/GR00T-N1.7-LIBERO (libero_10)
• Trainable: everything except the frozen VLM backbone (nvidia/Cosmos-Reason2-2B) — projector I/O layers, VL self-attention refinement, and the diffusion (DiT) action head (1.62B / 3.14B params, 51.5%)
• 8x GPU, DeepSpeed ZeRO-2
• Global batch size: 192
• 20 epochs (21,767 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: 0.1723 (from ~1.35-1.38 at start). Training took 10h42m on 8x24GB GPUs.