MODEL ROLE
A compact VLA can still require a careful robot interface.
SmolVLA uses a vision-language backbone and a specialized continuous-action expert. It can condition action generation on image observations, language instructions and robot state, but the checkpoint’s expected feature schema and action space still need to match the deployed embodiment.
The existing SO-101 teleoperation and no-video dataset smoke test do not establish that a SmolVLA policy has trained, been loaded or autonomously controlled our follower.
EVIDENCE REQUIREMENTS
Before training: observation/action semantics must be explicit.
A useful training dataset needs named camera streams, consistent joint/state ordering, action features, aligned timestamps, task labels and episode splits. Calibration digest and exact LeRobot/model revisions belong in the data card.
Our first dataset smoke run captured no video. It is therefore not presented as the multimodal training set required to substantiate a vision-conditioned SmolVLA task result.
Camera + proprioception
Pin image dimensions, frame names, timestamps and robot-state feature order.
Continuous command space
Validate units, joint ordering, gripper semantics and control mode against the actual follower.
Training/validation split
Record dataset revision, episode identities, held-out task starts and any excluded trials.
Checkpoint and inference
Pin policy type, config, model revision, action horizon and compute device.
FIRST REAL EXPERIMENT
Prefer a tiny, reproducible task before broad autonomy claims.
Start with a lightweight object-transfer task, supervised inside a bounded workspace, with a predeclared success condition and known safe-stop behavior. Evaluate the same task under a pinned checkpoint and held-out starts.
For the comparison phase, capture ACT and SmolVLA under matched trial conditions. For async inference, retain the same SmolVLA checkpoint so queueing and action-age effects can be distinguished from model changes.
RUNTIME BOUNDARY
A trained policy still is not a safety controller.
Successful action prediction does not prove calibration, collision avoidance, physical bounds or full runtime safety. OpenRGD treats VLA output as a proposal relative to the target body; independently governed runtime and adapter components own execution.
The public evaluation protocol exists today. The measured SmolVLA SO-101 result will remain empty until an actual run meets its evidence gate.