Yifei Yao

InfCoiL: Coordinated Planner-Controller Learning for Closed-Loop Physics-Based Human-Object Interaction

NeurIPS 2026 2026 · Accepted

Yude Zou^*, Yifei Yao^*, Yiwei Hao^*, Hanqing Wang, Xing Gao

* Equal contribution (co-first authors).

A flow-matching planner and a multi-morphology controller learn together through closed-loop reinforcement learning for physical human-object interaction.

InfCoiL planner-controller framework for human-object interaction — InfCoiL overview. Figure from coauthor Xing Gao’s research page.

InfCoiL connects text- and goal-conditioned motion planning with physical execution. A morphology-aware planner and unified controller operate in an autoregressive plan-and-execute loop. Coordinated reinforcement learning, including Semantic Flow-GRPO for the planner, addresses the distribution shift between generated plans and physical rollouts.

Accepted to NeurIPS 2026. The conference page and OpenReview record are linked above; no public code release is claimed here.

Reference

BibTeX

@inproceedings{yao2026infcoil,
  title = {InfCoiL: Coordinated Planner-Controller Learning for Closed-Loop Physics-Based Human-Object Interaction},
  author = {Yude Zou and Yifei Yao and Yiwei Hao and Hanqing Wang and Xing Gao},
  year = {2026},
  url = {https://openreview.net/forum?id=RvZ2OxrgcE},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS 2026)},
  note = {Accepted}
}

Metadata checked October 27, 2026. Original source ↗

← All publications