Constrained Generative Models for Robotic Learning

Overview

Robotic systems must generate actions that satisfy hard physical constraints — joint limits, collision avoidance, contact dynamics — while remaining adaptive to novel environments. I study how generative models can serve as constrained policy learners, combining the expressiveness of diffusion and flow-based planners with the guarantees needed for safe real-world deployment.

This includes incorporating kinematic and dynamic constraints into diffusion-based planners, composing safety constraints with reward-driven objectives, and enabling generative planning to transfer across embodiments.