rubatotree/classify-blocks-2-1-smolvla

classify-blocks-2-1 SmolVLA

An experimental SmolVLA policy trained on rubatotree/classify-blocks-2-1, the 512-episode hardware-geometry revision of the synthetic single-block pick-place dataset. The model starts from lerobot/smolvla_base and uses five joint positions in degrees plus one gripper command in percent.

Scope

This checkpoint has not been evaluated on a physical arm. The source demonstrations carry smoke_only=true and training_eligible=false; training used an explicit experimental admission for simulation plumbing and trajectory review. No hardware success rate, contact fidelity, or closed-loop generalization claim is made.

The exact training steps, batch size, dataset hashes, and strict checkpoint reload are recorded in training_report.json beside the weights.

Interface

observation.images.front	(3, 480, 640) RGB front camera
observation.state	(6,) degrees for five joints plus gripper percent
action	(6,) absolute command in the same units at 15 Hz
language	the dataset task string

Usage

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy

policy = SmolVLAPolicy.from_pretrained("rubatotree/classify-blocks-2-1-smolvla")

Use the policy only with an explicit simulation or supervised trial setup until a hardware rollout has been measured.