ChunkTrust

Adapting Execution Horizons for Robot Policies with Action-Expert Evidence

Paper · Project Page · Code

Public model assets for ChunkTrust. The maintained implementation, installation instructions, and training and evaluation recipes are in the code repository.

Data and checkpoints

Training data for the simulation experiments come from RoboTwin 2.0 and RoboCasa GR1 Tabletop. The QHA protocols use RoboTwin 2.0 clean50 demonstrations. Demonstration datasets and preprocessed teacher caches are not included in this repository.

Checkpoint group	Step	Published assets
pi0 RoboTwin 2.0 task policies	30000	8 task-specific policies with normalization statistics
pi0.5 RoboTwin 2.0 task policies	20000	8 task-specific policies with normalization statistics
pi0.5 RoboTwin 2.0 multitask policy	30000	50-task policy with normalization statistics
pi0 QHA, eight-task augmentation	5000	Adapter and configuration
pi0.5 QHA, eight-task augmentation	5000	Adapter and configuration
pi0.5 QHA, six-task training and two held-out tasks	10000	Adapter and configuration

QHA adapters must be used with the matching frozen base policy and normalization settings. See the training recipes and model guide for the checkpoint-to-protocol mapping.

The pi0.5 RoboCasa GR1 policy is hosted separately on ModelScope.

Download

The asset catalog lists all published checkpoint files at pinned revisions. From a clone of the code repository:

curl -L https://huggingface.co/Niugan/ChunkTrust/resolve/main/code/assets.json -o assets.json
python scripts/download_asset.py qha_heldout --catalog assets.json --destination checkpoints/qha-heldout/10000

Other adapter keys are qha_pi0_8task_step5000 and qha_pi05_8task_step5000. For base policies, use the task or multitask model key in the catalog.

Repository contents

• checkpoints/pi0/: pi0 task-specific policies.
• checkpoints/pi05/: pi0.5 task-specific and multitask policies.
• checkpoints/qha_*/: the three QHA adapters.
• code/: public source snapshot and download catalog.
• media/: introduction video and cover image.

Download the public source snapshot, corresponding to GitHub commit 27cfab89a82295cc534e9dd641bdec7c85499797. The 4K introduction is also available on GitHub.

Base-model and dataset licenses apply independently of the code repository's MIT license.