APPL-assets

APPL: Simulation Datasets and Pretrained Policies

Release assets for Agent Priors-guided Policy Learning (APPL).

Paper · Code and reproduction guide · Project page

These archives contain the paper's Exp1 and final five-task Exp2 data, trained policies and detailed evaluation evidence. Agent+VLA and real-robot artifacts are outside this distribution.

Files

SHA256SUMS contains the SHA256 of all three archives. README.txt provides the matching source commit and installation commands. The two experiment archives contain 191,640 files and 222 trained checkpoints in total.

Download and install

Download the three archives, SHA256SUMS and README.txt into one directory. Use the specific release tag or commit from Files and versions for repeatable downloads. These are experiment archives installed by the APPL release tools; they are not a tabular dataset for datasets.load_dataset.

From that download directory, on Linux with Pixi 0.80.0 or newer:

sha256sum --check SHA256SUMS
tar -xzf appl-source.tar.gz
cd appl
pixi install --locked
pixi run --locked python -m appl_release fetch --bundle exp1 --source ../appl-exp1-assets.tar
pixi run --locked python -m appl_release fetch --bundle exp2 --source ../appl-exp2-assets.tar
pixi run --locked python -m appl_release verify --assets
pixi run --locked tables

The installer validates the archive and every extracted file. Exp1 uses the root Pixi environment; Exp2 uses environments/exp2/pixi.toml. Follow the experiment guides for simulator installation, GPU selection and evaluation. Recomputing the released tables needs no GPU or model API. New construction agent and runtime agent sessions require the user's own compatible API access.

Scope and interpretation

• Exp1: six MetaWorld tasks, 2/5/10/20 demonstrations, 144 trained systems and 14,400 formal hidden-test outcomes in the matching source release.
• Exp2: five ManiSkill tasks and 560 released episodes across motion, task and composition suites; 96 frozen evaluation cases in the source release.
• The checkpoints correspond to the reported runs. This asset release does not represent new training or evaluation runs.
• These are simulation research artifacts. Reported simulation performance does not establish physical-robot performance.

License and citation

APPL-authored materials are released under the project's MIT license. Third-party components retain their own licenses; see LICENSE and THIRD_PARTY_NOTICES.md in the source archive. Use the accompanying CITATION.cff when citing APPL.

If you use these datasets, checkpoints or results, please cite:

@misc{jiang2026appl,
  title = {Agent Priors-guided Policy Learning},
  author = {Jiang, Puming and Hu, Tianrun and Du, Haozhe and Li, Yibo and Xue, Zhiwei and Li, Xinhu and Soh, Harold},
  year = {2026},
  eprint = {2609.35690},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2609.35690}
}