smolvla_cube_in_box_v2 — current model

SmolVLA for the SO-101 arm, task "put the cube in the box". The root of this repository holds run4's weights (object positions 1 and 2), so that --policy.path=denys1912/smolvla_cube_in_box_v2 keeps working unchanged.

Each experiment lives in its own repository: LeRobot loads config.json from the repo root only and cannot read subfolders, so filing the runs into folders inside one repository would make them unloadable.

run	data	best checkpoint	val loss
run5	166 episodes, four positions, stratified split	step 1500	0.17636
run4 ← the weights in root	86 episodes, positions 1–2, stratified split	step 1250	0.16032
run2	47 episodes, position 1 only	step 750	0.15922

The val losses are not comparable across runs: each holds out a different set of episodes (7 at one position, 13 across two, 25 across four). A higher number for a later run reflects a broader validation set, not a worse model.

Training code: github.com/denysk19/vla_training — a custom loop over LeRobot components with the held-out validation that lerobot-train does not have at all.