Every Embodied ACT Protected Diagnostic Checkpoint
This is the current ACT protection candidate for the Datawhale Every Embodied AMD ROCm tutorial.
Verified result
- Strict physical success:
15/30
- Three fixed seed groups of 10:
3/10 + 4/10 + 8/10
- Action dimension: 16-D
- Recipe: reset-aligned data, timestamp/object-init state, chunk size 20, 10 action steps, low-learning-rate continuation, and weighted recovery data
- Model SHA256:
b9b178377995a674a06bc5d1500c8e7e7fc5d02649268855f892b3987bf5bfeb4
This is a teaching and diagnosis checkpoint, not a claim that the historical 17/30 summary has been fully reproduced. The current score came from an AMD395 wrapper continuation; the matching native training recipe is included in the tutorial Notebook and should be run in a Jupyter-capable environment before calling it a Notebook-native score.
Tutorial
Contents
The weights/ directory contains the ACT model and configuration. Evaluation metadata and the recipe are kept in evaluation/. Raw datasets are intentionally omitted.
License and attribution
Please follow the licenses of ACT, the base implementation, and the Every Embodied project. This model card describes a reproducible tutorial artifact, not a new policy architecture.
