Document ChunkTrust models, reproduction status, and 4K video
Browse files
README.md
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- en
|
| 4 |
+
tags:
|
| 5 |
+
- robotics
|
| 6 |
+
- vision-language-action
|
| 7 |
+
- action-chunking
|
| 8 |
+
- chunktrust
|
| 9 |
+
library_name: openpi
|
| 10 |
+
---
|
| 11 |
+
# ChunkTrust
|
| 12 |
+
|
| 13 |
+
**Adapting Execution Horizons for Robot Policies with Action-Expert Evidence**
|
| 14 |
+
|
| 15 |
+
[GitHub](https://github.com/hf618/ChunkTrust) 路 [Source archive](code/ChunkTrust-source-0.1.0.tar.gz) 路 [Project page](https://hf618.github.io/ChunkTrust.github.io/)
|
| 16 |
+
|
| 17 |
+
<video controls src="https://huggingface.co/Niugan/ChunkTrust/resolve/main/media/ChunkTrust_work_intro_4k.mp4" width="100%"></video>
|
| 18 |
+
|
| 19 |
+
Full introduction: **4K, 30 fps, 2 min 57.5 s, English narration and subtitles**.
|
| 20 |
+
[Download the 4K MP4](media/ChunkTrust_work_intro_4k.mp4).
|
| 21 |
+
|
| 22 |
+
## Contents
|
| 23 |
+
|
| 24 |
+
- `checkpoints/qha_pi05_heldout6_step10000`: identified QHA-only head and assets.
|
| 25 |
+
- `checkpoints/pi05/...`: task-specific and multitask pi0.5 inference checkpoints as uploads complete.
|
| 26 |
+
- `checkpoints/pi0/...`: task-specific pi0 inference checkpoints as uploads complete.
|
| 27 |
+
- `media`: 4K introduction and cover.
|
| 28 |
+
|
| 29 |
+
The QHA-only head must be paired with the frozen task-specific base policy;
|
| 30 |
+
it is not a standalone VLA. See the GitHub asset catalog for complete, verified
|
| 31 |
+
uploads and immutable revisions. Do not use incomplete directories during upload.
|
| 32 |
+
|
| 33 |
+
QHA is trained on six RoboTwin tasks, with Blocks Ranking RGB and Place Bread
|
| 34 |
+
Basket held out from QHA training. The base policies remain task-specific.
|
| 35 |
+
Normalization, candidate sets and source-code versions must match the evaluation.
|
| 36 |
+
|
| 37 |
+
## Reproduction status
|
| 38 |
+
|
| 39 |
+
The code repository retains the original episode outcomes and records two
|
| 40 |
+
unresolved differences between available result records and the manuscript.
|
| 41 |
+
See [the reproduction audit](https://github.com/hf618/ChunkTrust/blob/main/docs/reproducibility.md).
|
| 42 |
+
This release does not claim a fresh full-suite rerun.
|
| 43 |
+
|
| 44 |
+
Base-model and dataset terms continue to apply. See the source repository's
|
| 45 |
+
third-party notices; its MIT code license does not relicense upstream weights.
|