Model Checkpoints for ManiSkill-HAB

RL (SAC, PPO) and IL (BC) baselines for ManiSkill-HAB. Each checkpoint includes a torch checkpoint policy.pt (model, optimizer/scheduler state, other trainable parameters) and a train config config.yml with hyperparemeters and env kwargs.

RL Pick/Place policies are trained using SAC due to improved performance, while Open/Close is trained with PPO for wall-time efficiency (see Appendix A.4.3). All-object RL policies are under all/ directories, while per-object policies are under directories labeled by the object name. IL policies do not have per-object Pick/Place variants.

To download these policies, run the following:

huggingface-cli download arth-shukla/mshab_checkpoints --local-dir mshab_checkpoints

If you use ManiSkill-HAB in your work, please consider citing the following:

@article{shukla2024maniskillhab,
    author		 = {Arth Shukla and Stone Tao and Hao Su},
    title        = {ManiSkill-HAB: A Benchmark for Low-Level Manipulation in Home Rearrangement Tasks},
    journal      = {CoRR},
    volume       = {abs/2412.13211},
    year         = {2024},
    url          = {https://doi.org/10.48550/arXiv.2412.13211},
    doi          = {10.48550/ARXIV.2412.13211},
    eprinttype   = {arXiv},
    eprint       = {2412.13211},
    timestamp    = {Mon, 09 Dec 2024 01:29:24 +0100},
    biburl       = {https://dblp.org/rec/journals/corr/abs-2412-13211.bib},
    bibsource    = {dblp computer science bibliography, https://dblp.org}
}