1. Speaker:
Daehyung Park (School of Computing, KAIST)

2. Title:
Robot May Not Work as We Want: Pathways to Precise, Constraint-Aware Behavior
Title partially revised.

3. Date & Time:
Thursday, May 21, 2026, 4:00 PM–5:30 PM

4. Venue:
L409

5. Abstract:
While recent advances in robot foundation models and vision-language-action models have accelerated the development of general-purpose robots, these systems still struggle to satisfy the levels of precision and safety required for reliable real-world deployment. In particular, current robot policies often fail to consistently comply with diverse safety constraints while generating the accurate and fine-grained behaviors demanded in practical environments.

In this talk, I will present our recent efforts to address these challenges from two perspectives. First, I will introduce an SSM-based Policy (SPo), which enables multi-granularity learning and action generation for precise robotic manipulation while overcoming the limitations of transformer-based policies. Second, I will discuss constraint learning methods that infer safety and operational constraints from human demonstrations and transfer them to unseen environments and tasks.

Together, these advances highlight a pathway toward robotic systems that move beyond general versatility to become precise, safe, and trustworthy partners for real-world industrial and collaborative applications.