makermods/smolvla_200ep_blue_cube_orange_box

SmolVLA · blue cube → orange box (SO-101)

SmolVLA fine-tuned from lerobot/smolvla_base on makermods/200ep_blue_cube_orange_box (199 episodes, 55,897 frames @ 30 fps, SO-101 follower, front + wrist cameras).

Training

• 20,000 steps, batch size 64 (~23 epochs), bf16 AMP, RTX 4090
• lr 1e-4, cosine decay to 2.5e-6 over 20k steps, 1k warmup
• Frozen vision encoder, action expert only (train_expert_only=true)
• Final smoothed loss ≈ 0.045
• Full config in train_config.json

Inference note

Trained with --rename_map='{"observation.images.front": "observation.images.camera1", "observation.images.wrist": "observation.images.camera2"}' — the policy expects camera1 (front) and camera2 (wrist) keys at inference; apply the same rename map when deploying.