Mehul2112/smolvla_black_cube_expert

SmolVLA — so100_black_cube (expert-only fine-tune)

Fine-tuned lerobot/smolvla_base on Mehul2112/so100_black_cube_merged (SO-100, single front camera, 130 episodes).

Recipe

• Mode: expert-only — VLM backbone frozen, action expert trained (train_expert_only=true, freeze_vision_encoder=true).
• Learnable params: ~100M of ~450M (action expert only).
• Batch size: 64 · AMP: bf16 · Optimizer LR: ~1e-4.
• Hardware: single RTX 4090 (24 GB).

Checkpoint

This repo holds the step-6000 checkpoint.

Step	Train loss	Eval loss (15% held-out)
2000	0.179	0.319 ← eval minimum
4000	0.121	0.348
6000	0.090	0.418

Note: Eval loss reached its minimum at step 2000 and rose afterward (overfitting on a small 110-episode training split). The step-6000 checkpoint here has the lowest training loss but is past the generalization optimum; for best real-robot generalization, the step-2000 checkpoint is preferable.

Usage

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
policy = SmolVLAPolicy.from_pretrained("Mehul2112/smolvla_black_cube_expert")