We have 1 paper accepted by NeurIPS 2026!

We have one paper accepted by NeurIPS 2026! This work introduces structured multimodal skill learning on the unit hypersphere for robust robot manipulation. Detailed information is provided below.

---

HyperSkill: Multi-Modal Skill Learning on the Unit Hypersphere

Authors: Erdemt Bao, Jun Chen, Weijun Qin, Shaopeng Li, Ming Li, Mengchen Zhao, Ziqian Zeng, Cen Chen, Huiping Zhuang

Abstract:

Multi-modal robot trajectories often contain several valid ways to accomplish the same task, yet existing skill representations can suffer from mode collapse, loss of multimodal structure, and a mismatch between planning and execution. HyperSkill represents structured skills on the unit hypersphere and uses a von Mises-Fisher mixture prior to capture diverse skill modes. It further introduces Geodesic Skill Alignment to align predicted prior skills with posterior skills inferred from demonstrations. Experiments on the LOReL Sawyer and Franka Kitchen benchmarks show strong performance in both seen and unseen settings, together with improved robustness to skill perturbations.

Paper: [To be added]

---