moonbot SmolVLA — three blocks stack, WITH force/torque
SmolVLA fine-tuned from lerobot/smolvla_base (full fine-tune: vision encoder and VLM trainable, as in the pi0 comparison runs; cameras renamed front->camera1, eef->camera2, right->camera3) on
gdiazsrl/three_blocks_stack_sep22.
Comparison run. One of six runs comparing pi0 / SmolVLA / ACT with vs without force/torque. Same dataset, batch 16, 20,000 steps and seed as the pi0 run drashutoshspace/moonbot_pi0_three_blocks_stack; the force/torque twin of this model is drashutoshspace/moonbot_smolvla_no_tf.
Training curves: https://wandb.ai/drmishra-space/lerobot/runs/rn1je6y9
Deployment contract: 21-dim state = joint_read (8) + tip_pos (7) + F_ee (6) force/torque; 3 cameras; 8-dim action.
Checkpoints
checkpoint-005000, -010000, -015000, -020000 (same steps as the counterpart run).
Each folder contains weights, pre/post-processors, this run's rosetta contract and DEPLOY.md.
The final checkpoint also includes training_state/. Complete checkpoints with optimizer state are
also on the team OneDrive under DATA/Processed/smolvla_tf/.
Training
| batch size | 16 |
| steps | 20,000 |
| optimizer / schedule | the algorithm's own LeRobot preset |
| seed | 1000 (default) |
| validation split | none (as in the pi0 runs) |