From Noise to Control: Parameterized Diffusion Policies

We propose Parameterized Diffusion Policy (PDP), a framework that learns a diffusion policy parameterized in a smooth continuous space. By structuring a latent manifold such that distances between latents' values reflect the semantic similarity of physical trajectories, we transform diffusion from a mechanism of stochastic diversity into a precise tool for behavior steering. Our approach also enables smooth interpolation between known strategies and efficient generalization to novel constraints without the need to update policy weights. We demonstrate that PDP significantly improves adaptation performance on complex multimodal benchmarks in both simulation and real-robot hardware compared to regular diffusion policy, particularly in scenarios requiring the discovery of novel behaviors.

Robots often need to solve the same task in different ways. For example, a robot may need to open a drawer, pick up a cup, or move around obstacles, but the best motion can change when an obstacle blocks its usual path. Modern diffusion-based robot policies can generate many possible motions, but it is often hard to control which motion they choose, especially when the environment changes. In this work, we introduce Parameterized Diffusion Policies, a method that gives the robot a more organized way to choose and adjust its behavior. The key idea is to learn a compact behavior space where similar robot trajectories are placed close together and different strategies are placed farther apart. This lets the policy adapt by moving within this space, rather than relying on random sampling or retraining the whole model. In simulation and real-robot experiments, this approach helps robots more reliably select feasible behaviors and adapt to new constraints from limited demonstration data.