Safe Diffusion Policies for Robot Learning
Project Summary
Diffusion policies have shown strong performance on complex, multimodal robot control tasks, but they offer no inherent guarantees against unsafe behavior at inference time — a critical gap for long-horizon decision-making in safety-sensitive settings. This project investigates inference-time safety mechanisms for diffusion policies, aiming to shape or filter the denoising process itself so that generated action sequences respect safety constraints without retraining the underlying policy.
To study this, I built a PyTorch training and evaluation framework for Diffusion Transformer policies spanning both navigation and manipulation domains, enabling controlled experiments on the tradeoff between safety intervention and task performance. Early experiments on path-planning tasks show promising results, while receding-horizon control and manipulation settings surface open challenges around closed-loop consistency when safety interventions interact with the policy’s own multi-step denoising dynamics.
Principal Researcher (Mar 2026 — Present)
- Investigating inference-time safety mechanisms for diffusion policies in long-horizon robot decision-making.
- Built a PyTorch training and evaluation framework for Diffusion Transformer policies across navigation and manipulation domains.
- Designed experiments on safety-performance tradeoffs and closed-loop consistency, demonstrating promising path-planning results and identifying challenges in receding-horizon control and manipulation.
Tech: PyTorch, Diffusion Transformers, robot learning, path planning, Gymnasium.
