MasterYip/PredActor_Artifacts

PredActor Evaluation Artifacts

Official evaluation checkpoints for PredActor, a predictive action diffusion controller for steerable Unitree G1 locomotion. The model jointly predicts future states and actions, uses its internal state trajectory for guidance, and directly executes the selected 29-DoF action.

Overview

This repository contains the two learned artifacts required by the public MuJoCo evaluation path:

• the released PDP051 text-conditioned PredActor policy;
• the fine-tuned G1 MotionCLIP encoder used for semantic conditioning.

Source code, the browser-based evaluator, and G1 robot assets are maintained in the public PredActor repository. Training datasets, experiment outputs, hardware deployment bundles, and additional checkpoints are not included here.

Quick Evaluation

Install uv, then run:

git clone https://github.com/MasterYip/PredActor.git
cd PredActor
uv sync --locked
uv run --locked python scripts/hf_download.py --filter checkpoints
uv run --locked predactor-eval

The downloader recreates the expected layout under the Git-ignored Artifacts/ directory and verifies the published byte sizes and SHA-256 identities. The evaluator opens http://127.0.0.1:8765/, uses CUDA when available, and otherwise falls back to CPU.

Released Artifacts

Artifact	Role	Size	SHA-256
checkpoints/predactor/pdp051/latest.ckpt	Text-conditioned PredActor policy	49,212,894 bytes	2d963b32786f2989c6472726df9fcfe6b385590127e12e1f549a4b7d77488b2e
checkpoints/motionclip/g1-model-xyz-clip/checkpoint_0100.pth.tar	Fine-tuned G1 MotionCLIP encoder	542,749,069 bytes	66a127df4958b346089b2020f2705c7456d9db0ee8b4bd9518608b708b35fc3c

SHA256SUMS records the same checkpoint identities in a format accepted by sha256sum --check.

Repository Layout

PredActor_Artifacts/
|-- README.md
|-- SHA256SUMS
|-- assets/
|   |-- predactor-readme-banner.svg
|   |-- predactor-preview.png
|   `-- institution-strip.svg
|-- checkpoints/
|   |-- predactor/pdp051/latest.ckpt
|   `-- motionclip/g1-model-xyz-clip/checkpoint_0100.pth.tar
`-- dataset/
    `-- README.md

dataset/ is reserved for future public datasets and currently contains no dataset payloads.

Model Details

• Policy: PDP051 using the public g1prdp_cond_diffuse evaluation profile.
• Robot: Unitree G1 with 29 actuated degrees of freedom.
• Inputs: proprioceptive observation history, a 512-dimensional semantic condition, and optional whole-body guidance commands.
• Outputs: a denoised action trajectory from which the current joint action is selected for direct execution.
• State prediction: future states remain internal to the policy and provide a differentiable target for classifier guidance.
• Public runtime: local MuJoCo evaluation through the PredActor Web UI.

Intended Use

These artifacts support research reproduction and local simulation evaluation of the released PredActor controller. They are intended for use with the hash-pinned public code and configuration. The MotionCLIP checkpoint is the semantic encoder expected by the released policy and should not be substituted without revalidating compatibility.

Limitations and Safety

• The public release is an evaluation package, not a training distribution.
• MuJoCo startup demonstrates software and checkpoint compatibility; it does not establish hardware safety or guarantee locomotion quality in a new environment.
• Real-robot deployment requires separate control, safety, and hardware validation that is outside this repository.
• PyTorch checkpoints use pickle-compatible deserialization. Load only files whose hashes match the published values and only from trusted sources.
• Performance may change with simulator, driver, GPU, or dependency versions outside the locked public environment.

Evaluation and Demos

The PredActor project page contains the method overview and simulation and hardware demonstrations. The public code repository provides the reproducible MuJoCo evaluation workflow.

License

PredActor is released under the MIT License. Third-party dependencies and robot assets remain subject to their respective upstream terms.

Citation

@misc{ye2026predactorpredictiveactiondiffusion,
  title={PredActor: Predictive Action Diffusion for Steerable Onboard Humanoid Control},
  author={Lei Ye and Haibo Gao and Yitang Li and Peng Xu and Zetong Jing and Junhan Sun and Fanrong Dong and Ziqi Han and Xue Wang and Jianhua Sun and Cewu Lu and Hao Zhao and Liang Ding},
  year={2026},
  eprint={2609.24840},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2609.24840},
}