gr00t-libero-x-level123

GR00T N1.7 — LIBERO-X Level 1-3 Extension Finetune

Finetune of nvidia/GR00T-N1.7-LIBERO (libero_10 checkpoint) on a 152-task subset of meituan/LIBERO-X, covering the full requested Level 1, Level 2, and Level 3 EXTENSION_KITCHEN/EXTENSION_STUDY task lists (218 requested scene/task pairs, collapsing to 152 distinct instruction strings due to LIBERO-X's reuse of identical phrasing across multiple scenes).

Training data

• Source: meituan/LIBERO-X (LeRobot v2.1 format), filtered by exact task-description match against the requested .bddl filenames (resolved via each file's :language field from meituan/LIBERO-X on GitHub).
• 1,076 episodes, 371,555 frames, 152 tasks.
• Embodiment: libero_sim (built-in EmbodimentTag.LIBERO_PANDA, panda arm, 8-dim state / 7-dim action, no custom modality config needed).

Training config

• Base checkpoint: GR00T-N1.7-LIBERO/libero_10
• 7x NVIDIA RTX A5000 (one GPU held another user's job, avoided)
• Global batch size 168 (24/GPU), learning rate 3e-5, weight decay 1e-5, warmup ratio 0.05, cosine schedule
• tune_projector=True, tune_diffusion_model=True, tune_vlln=True (inherited from base checkpoint), tune_llm=False, tune_visual=False
• DeepSpeed ZeRO-2, bf16
• 20 epochs, 44,233 steps, ~24h 43m wall time
• Final train loss: 0.167

Contents

Weights-only checkpoint (matches the base GR00T-N1.7-LIBERO release format): model-*.safetensors + model.safetensors.index.json, config.json, processor_config.json, embodiment_id.json, statistics.json. DeepSpeed optimizer/RNG state is excluded (train-only, not needed for inference or further LoRA-style finetuning from this checkpoint).