ChunkTrust

Adapting Execution Horizons for Robot Policies with Action-Expert Evidence

GitHub · Source archive · Project page

Full introduction: 4K, 30 fps, 2 min 57.5 s, English narration and subtitles. Download the 4K MP4.

Contents

• checkpoints/qha_pi05_heldout6_step10000: identified QHA-only head and assets.
• checkpoints/pi05/...: task-specific and multitask pi0.5 inference checkpoints as uploads complete.
• checkpoints/pi0/...: task-specific pi0 inference checkpoints as uploads complete.
• media: 4K introduction and cover.

The QHA-only head must be paired with the frozen task-specific base policy; it is not a standalone VLA. See the GitHub asset catalog for complete, verified uploads and immutable revisions. Do not use incomplete directories during upload.

QHA is trained on six RoboTwin tasks, with Blocks Ranking RGB and Place Bread Basket held out from QHA training. The base policies remain task-specific. Normalization, candidate sets and source-code versions must match the evaluation.

Reproduction status

The code repository retains the original episode outcomes and records two unresolved differences between available result records and the manuscript. See the reproduction audit. This release does not claim a fresh full-suite rerun.

Base-model and dataset terms continue to apply. See the source repository's third-party notices; its MIT code license does not relicense upstream weights.