Fine-tuned from lerobot/smolvla_base @ d9f33c94a60fb382c90dea2164c96845bd955e28 with LeRobot 0.6.1 on
1,076 domain-randomized MuJoCo demonstrations (173,683 frames). Vision encoder and language model frozen;
action expert and state/action projections trained (20k steps, batch 16, peak 4.62 GiB under a 7 GiB cap).
Inputs: observation.state (5 joints in degrees + gripper 0-100), observation.images.wrist / overhead
(480x640 RGB; renamed to camera2 / camera1 by the saved preprocessor), task "Insert the peg into the hole."
Output: absolute joint targets in the same units. Trained on rendered images only.
Code, evaluation and report: the accompanying repository (policy.py, evaluate.py).
pretrained_model/model.safetensors 907 MBpretrained_model/policy_postprocessor_step_0_unnormalizer_processor.safetensors 7.8 kBpretrained_model/policy_preprocessor_step_5_normalizer_processor.safetensors 7.8 kBlerobot/smolvla_base