library_name: lerobot tags:
- robotics
- embodied-ai
- smolvla
- rocm
- amd
Every Embodied SmolVLA ROCm PnP
This is the protected SmolVLA checkpoint used by the Datawhale Every Embodied AMD ROCm tutorial.
Verified result
- Strict physical success:
57/60 - Red instruction:
27/30 - Blue instruction:
30/30 - Task: MuJoCo mug-to-plate pick and place
- Training recipe: weighted blue sampling, selected checkpoint at step 500
- Hardware path: AMD ROCm teaching environment
- Model SHA256:
9e372a2babf79ec9530ecf75719cb62f02974485278751333867cedb5aa744fa
The score is reported with the tutorial’s strict physical predicate. It is not a zero-shot claim: this is a fine-tuned checkpoint and must be loaded with the matching environment, action normalization, camera order, and instruction format.
Tutorial
See the Every Embodied AMD ROCm topic and 14_smolvla_end_to_end.ipynb.
Contents
The weights/ directory contains model and configuration files. Evaluation metadata is kept in evaluation/. Optimizer state and raw datasets are intentionally omitted.
License and attribution
Please follow the licenses of SmolVLA, LeRobot, the base checkpoint, and the Every Embodied project. This model card describes a reproducible tutorial artifact, not a new foundation model.