SmolVLA Piper is a robotics policy checkpoint trained with Hugging Face’s LeRobot tooling. Under the hood it’s based on lerobot/smolvla_base, a compact vision-language-action (VLA) model designed to be efficient enough for consumer-grade hardware. This particular checkpoint is trained on a Piper arm dataset (ISdept/piper_arm), so the expectation is that it can map observations to control actions in that environment.
The model card is intentionally minimal and points readers at the SmolVLA paper and the LeRobot docs for the full training/inference workflow. That’s actually useful if you’re evaluating VLA policies: it makes it clear that this repo is primarily a runnable artifact (weights + config) rather than a long narrative.
What to try first: if you already have a LeRobot setup, run a short eval recording session (for example, lerobot-record) with a few episodes and inspect failure modes. If you don’t, treat this as a template repo for how to publish a robotics policy checkpoint (including license and metadata) so downstream users can load it with standard transformers/LeRobot tooling.
Source listing: https://huggingface.co/models?sort=modified