Title: From Noise to Control: Parameterized Diffusion Policies

Abstract:

This work introduces Parameterized Diffusion Policy (PDP), a novel framework designed to learn diffusion policies that are conditioned on low-dimensional, continuous parameters situated within a learned behavior manifold. By engineering this manifold such that the distances between latent representations correspond to the semantic similarity of physical trajectories, we repurpose diffusion from a mere source of stochastic diversity into a precise, optimizable instrument for steering behavior. This methodology facilitates smooth interpolation between established strategies and allows for efficient adaptation to new constraints without the need to update policy weights. Our experiments, conducted on both simulated environments and real-robot systems using complex multimodal benchmarks, demonstrate that PDP substantially outperforms standard diffusion policies. This improvement is particularly notable in scenarios that demand the synthesis of novel behaviors.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC