Modelpublic

omkarpatil/ffw_sg2_put-blocks-in-bowl_smolvla

sourceHugging Faceapache-2.0updated 29d agoView on Hugging Face

0likes30downloads

Model Card

SmolVLA — FFW SG2 rev1 — "Pick the two blocks up and put them in the green bowl"

lerobot/smolvla_base fine-tuned with LeRobot 0.6.1 (cyclo_intelligence submodule) on 78 teleop demos (omkarpatil/put-blocks-in-bowl, 15 fps). Checkpoint: step 20 000 of 60 000, batch 64, seed 1000, no image augmentation.


Cameras	`cam_left_head`, `cam_left_wrist`, `cam_right_wrist` (rename map → `camera1..3`, stored in `policy_preprocessor.json`)
State	22-D: arml×7, gripperl, armr×7, gripperr, head×2, lift, base vel×3
Action	16-D absolute joint targets: arml×7, gripperl, armr×7, gripperr — chunk 50 @ 15 Hz
Task	`Pick the two blocks up and put them in the green bowl`

Note: config.json input_features still carries the base model's shapes (state 6, 256×256) — LeRobot does not refresh them on a --policy.path fine-tune. The model normalizes with the 22-D dataset stats in the saved preprocessor. The cyclo_intelligence LeRobot engine on branch tabletop-manip (omkarpatil18 fork) handles this; head/lift/base are not commanded by this policy.