- Speaker : Prof. Daehyung Park
- Date : May 21, 2026
- Affiliation : KAIST School of Computing
- Category : Invited Seminar
Robot may not Work as We Want: Pathways to Precise, Constraint-Aware Behavior
Abstract
While AI and data-driven methods have significantly advanced robot capabilities, deploying robots in the real world remains challenging. This seminar introduces innovative frameworks designed to achieve precise, constraint-aware robotic behaviors. Prof. Park discusses key methodologies including Proactive Constraint Learning utilizing Control Barrier Functions to preemptively limit actions, and Transferable Constraint Learning to extract safety rules from expert demonstrations. The session further explores dynamic navigation context-awareness and diffusion-based policies for handling complex real-world variables.
Presentation Overview
This presentation covers the following key topics:
- Proactive and Transferable Constraint Learning to proactively limit behaviors and apply safety-residual rewards to new environments.
- SuReNav (Superpixel Graph-based Constraint Relaxation) for intelligent safety-efficiency trade-offs during dynamic robot navigation.
- Action Resolution Learning (DiSPo) leveraging diffusion-based policies to enhance representation power in multi-granularity tasks.
