RAIN v0.1.1 — public ManiSkill release

Public ManiSkill implementation of RAIN: training, calibration and pretrained evaluation for eight tasks, with project-page and manuscript links.

• Organized code, installation/evaluation/training guides, software citation, contribution guidance and CI.
• Public model and reference simulator downloads use Python's standard library; no GitHub login is needed. Archive sizes, SHA-256 digests and member lists are verified before use.
• Pretrained model bundles and the reference ManiSkill source remain at v0.1.0, with their original bytes and checksums.
• Attached RAIN_arXiv.pdf is the 9-page author-identified, two-column manuscript. Its SHA-256 is ccdabda89b73b0306667e09f8007d86ecb0666acfb3130166847170d03679655.

Validation: 21 CPU regression/packaging checks passed. Prior GPU integration checks cover one episode per task; full retraining and the archived 800-episode experiment were not repeated for this release.

Project page: https://rain-policy.github.io/

Code, documentation and original RAIN model weights use MIT. The manuscript is distributed separately; third-party dependencies, simulator assets and datasets retain their own terms. The paper is published as arXiv:2610.19765; the repository includes BibTeX and GitHub citation metadata.

Official published PDF: https://arxiv.org/pdf/2610.19765

RAIN v0.1.0 — pretrained models

Pretrained RAIN models for eight ManiSkill tasks, used by the public code at https://github.com/rain-policy/RAIN.

• rain-models-v0.1.0.tar.gz: 32 portable model/configuration files for eight tasks. SHA-256: 0304986672ca52b43ca170316e36886b39126ac577f64af694fba8c9afe114ff.
• rain-maniskill-source-fb3328d.tar.gz: the retained reference ManiSkill source, with upstream licenses and asset notices. SHA-256: 75350b7cab59c60369a32910485c3d612f92339d5f09be8c67145da575af8e87.
• RAIN-MODELS-LICENSE.txt: MIT license for the original RAIN model weights.

The original model and simulator archives are unchanged. Model path metadata is sanitized; all 3,104 serialized tensors and numeric routing parameters match the archived originals. Each task passed a one-episode GPU integration check against its corresponding archived episode. Full retraining and the 800-episode experiment were not repeated during packaging.

Use the current repository or the v0.1.1 code release, then run python tools/download_models.py. Downloads require no GitHub login. Installation, evaluation and training guides are linked in the README.

Project page: https://rain-policy.github.io/

Original RAIN code, documentation and these eight-task model weights use MIT. The separate ManiSkill source, simulator assets and demonstration datasets retain their upstream terms.

Paper: arXiv:2610.19765 · PDF.