Sampling-Based Constrained Policy Optimization

Principal Researcher (Sep 2025 — Dec 2025)

• Proposed a sampling-based weight-space projection framework for enforcing safety constraints during policy updates, enabling scalable safe learning for large neural policies.
• Formulated policy optimization as a convex projection in parameter space, with theoretical guarantees that projected updates preserve objective improvement.
• Derived a sufficient stability condition linking weight updates to closed-loop constraint satisfaction, and embedded it directly into the optimization pipeline.
• Empirically demonstrated complete rejection of harmful supervision and safe performance gains in regression and imitation learning under adversarial experts.

Tech: PyTorch, Convex Optimization, Reinforcement Learning, Safe Learning.