smolvla-cupflow

SmolVLA (lerobot/smolvla_base) fine-tuned on jcapmeil/cupflow-lerobot (150 episodes, 604k frames).

Training

LeRobot	0.4.4
Steps	20,000 (batch 64, ~2.1 epochs)
Final loss	0.012 (step 100: 0.439, 10k: 0.016)
LR schedule	warmup 666 → cosine decay to 2.5e-06
Hardware	1× L40S, ~4 h
Trainable params	100M of 450M

Checkpoint: step 20,000 (final).

Camera names (important)

The model was trained with the dataset's cameras renamed:

dataset key	policy key
observation.images.top	observation.images.camera1
observation.images.wrist	observation.images.camera2

At inference, either name the robot's cameras camera1 and camera2, or apply the same rename map. Otherwise the policy will either fail or receive the wrong images. The config also lists camera3, inherited from smolvla_base; it was not present in training.

Usage

--policy.path=jcapmeil/smolvla-cupflow

Training flags: --rename_map='{"observation.images.top": "observation.images.camera1", "observation.images.wrist": "observation.images.camera2"}' --tolerance_s=1e-3