PACT Data and Checkpoints

This repository hosts the data and pretrained policy checkpoints for PACT: Self-Evolving Physical Safety Alignment for Diffusion Policies in Embodied Manipulation.

PACT is a self-evolving post-training framework for aligning pretrained diffusion policies with physical safety constraints in embodied manipulation. It uses self-rollouts and automatically computed physical constraints to distill constraint gradients into diffusion policies, improving safety without requiring demonstrations, task rewards, interventions, or outcome annotations.

Website - GitHub - Paper

---

Highlights

• Self-evolving: Aligns diffusion policies from self-rollouts without demonstrations, interventions, rewards, or outcome annotations.
• Efficient optimization: Distills constraint gradients into the policy, providing dense supervision across diffusion timesteps.
• Curriculum alignment: Progressively tightens constraints to preserve task competence while improving safety.
• Foundation-model compatible: Plugs into diffusion-based policies, including flow policies, VLAs, and WAMs, without architecture modifications.
• Theoretical control: Bounds policy shift and supports monotone improvement.
• Proven in practice: Improves both task success and safety across simulation and real-world manipulation.

---

Repository Contents

This Hugging Face dataset repository contains file-based assets used by the official PACT codebase:

• Pretrained base policy checkpoints for Diffusion Policy on RoboTwin tasks.
• Pre-generated instruction datasets used during PACT post-training.
• Shared environment metadata required by the released instruction data.

The released task set includes:

• handover_apple
• handover_block
• pick_diverse_bottles
• pick_dual_bottles
• place_dual_shoes
• pour_water_to_cup
• stack_blocks_two

This repository is intended for file-based download and use with the official GitHub codebase. The Hugging Face dataset viewer is not the primary interface for these assets.

---

Quickstart

Install Git LFS if needed, then download the repository:

git lfs install
git clone https://huggingface.co/datasets/Ethan-pooh/pact

Alternatively, download with the Hugging Face CLI:

huggingface-cli download Ethan-pooh/pact \
  --repo-type dataset \
  --local-dir ./pact_hf

The instruction data is provided as data.tar.gz. Untar it and place the extracted files under the PACT repository:

mkdir -p /path/to/PACT/data
tar -xzf ./pact_hf/data.tar.gz -C /path/to/PACT/data

Please refer to the official PACT GitHub repository for installation, post-training, and evaluation commands.

---

Expected Directory Layout

After downloading, place the pretrained base policy checkpoints under the PACT repository as follows:

PACT/policy/DP/checkpoints/
├── handover_apple-demo_randomized-200-0/600.ckpt
├── handover_block-demo_randomized-200-0/600.ckpt
├── pick_diverse_bottles-demo_randomized-200-0/600.ckpt
├── pick_dual_bottles-demo_randomized-200-0/600.ckpt
├── place_dual_shoes-demo_randomized-200-0/600.ckpt
├── pour_water_to_cup-demo_randomized-200-0/600.ckpt
└── stack_blocks_two-demo_randomized-200-0/600.ckpt

After untarring data.tar.gz, the pre-generated instruction dataset and shared environment metadata should follow this structure:

PACT/data/
├── data/handover_apple/demo_randomized/instructions
├── data/handover_block/demo_randomized/instructions
├── data/pick_diverse_bottles/demo_randomized/instructions
├── data/pick_dual_bottles/demo_randomized/instructions
├── data/place_dual_shoes/demo_randomized/instructions
├── data/pour_water_to_cup/demo_randomized/instructions
├── data/stack_blocks_two/demo_randomized/instructions
└── env_meta.pkl

Once the files are arranged, evaluate a base policy with:

cd policy/DP
bash eval_dr.sh pick_dual_bottles demo_randomized demo_randomized 200 0 600 0

Run PACT post-training from the repository root with:

CUDA_VISIBLE_DEVICES=0,1,2,3 bash policy/DP/on_policy_distill_multigpu.sh pick_dual_bottles onpolicy_randomized 200 0 14

---

Intended Uses and Limitations

Intended uses

• Research on robot manipulation, diffusion policies, and physical safety alignment.
• Reproducing the PACT post-training pipeline on RoboTwin tasks.
• Evaluating pretrained base policies and post-trained PACT policies using the official codebase.

Limitations

• The checkpoints and instruction data are tied to the released RoboTwin task configurations.
• Post-training depends on correct installation of RoboTwin, Diffusion Policy, PACT cost functions, and task assets.
• Results may vary with simulator version, asset placement, random seeds, and the implemented physical constraint functions.

Safety and responsible use

• These assets are intended for simulation-based research in RoboTwin.
• Do not interpret simulated safety improvements as a direct guarantee of real-world robot safety without additional validation.

---

Changelog

• 2026-06: Initial release of PACT data and pretrained base policy checkpoints on Hugging Face.

---

Citation

If you find this dataset, checkpoints, or codebase useful, please cite:

@article{wu2026pact,
  title={PACT: Self-Evolving Physical Safety Alignment for Diffusion Policies in Embodied Manipulation},
  author={Wu, Lingxuan and Zhu, Zijian and Wang, Lizhong and Ying, Chengyang and Chen, Huayu and Yang, Xiao and Liu, Fangming and Zhu, Jun},
  journal={arXiv preprint arXiv:2606.08414},
  year={2026}
}

---

Contact

• Project page: https://ethan-iai.github.io/pact/
• Paper: https://arxiv.org/abs/2606.08414
• GitHub: https://github.com/thu-ml/PACT
• Hugging Face: https://huggingface.co/datasets/Ethan-pooh/pact