We are excited to share that two papers from CLeAR were accepted to R:SS 2026! More information about the papers is coming soon, but here’s a snapshot:
-
Guided Streaming Stochastic Interpolant Policy. Led by Puming (Oscar), this work introduces a principled inference-time guidance framework for streaming generative robot policies. By deriving the optimal guidance law for stochastic interpolants and combining it with streaming action generation, the method enables robots to reactively adapt during execution, such as avoiding moving obstacles or following user-specified grasping preferences. The framework supports both training-free guidance through STEG and training-based guidance through CCG. A pre-print is available here.
-
SkillVLA: Tackling Combinatorial Diversity in Dual-Arm Manipulation via Skill Reuse. This work studies how bimanual vision-language-action models can reuse learned single-arm skills and recompose them into new dual-arm behaviors. SkillVLA introduces a two-level reasoning framework that identifies skill structure, separates per-arm action generation when appropriate, and enables inter-arm communication when cooperation is needed. This allows the robot to generalize to unseen combinations of skills while maintaining strong performance on cooperative and long-horizon tasks. A pre-print is available here.
