Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising

TL;DR

Diffusion policies enable probabilistic action modeling for robotic manipulation but inference remains sequential and costly. The authors propose Genetic Diffusion Policy (GDP), a population-based denoising strategy that uses OoD-aware fitness to select in-distribution trajectories, reducing inference steps without retraining. Across 14 manipulation tasks, GDP achieves up to 20% improvements over standard diffusion policies and often matches or surpasses shortcut models, even with as few as two denoising steps. The study also shows that simple tweaks to the denoising schedule and noise scale can boost performance, emphasizing that embodied AI requires diffusion deployments tailored to low-dimensional action spaces rather than direct transfers from image generation.

Abstract

Diffusion models, such as diffusion policy, have achieved state-of-the-art results in robotic manipulation by imitating expert demonstrations. While diffusion models were originally developed for vision tasks like image and video generation, many of their inference strategies have been directly transferred to control domains without adaptation. In this work, we show that by tailoring the denoising process to the specific characteristics of embodied AI tasks -- particularly structured, low-dimensional nature of action distributions -- diffusion policies can operate effectively with as few as 5 neural function evaluations (NFE). Building on this insight, we propose a population-based sampling strategy, genetic denoising, which enhances both performance and stability by selecting denoising trajectories with low out-of-distribution risk. Our method solves challenging tasks with only 2 NFE while improving or matching performance. We evaluate our approach across 14 robotic manipulation tasks from D4RL and Robomimic, spanning multiple action horizons and inference budgets. In over 2 million evaluations, our method consistently outperforms standard diffusion-based policies, achieving up to 20\% performance gains with significantly fewer inference steps.

Figures (11)

• Figure 1: Comparison of normalized scores of Genetic Diffusion Policy to shortcut and diffusion policy baselines on Adroit Hand tasks.
• Figure 2: Genetic Denoising Process. One starts from pure Gaussian noisy samples, fitness scores are computed and used a weights for a multinomial selection. Selected samples are duplicated to replace deleted samples. Then, apply a denoising step following a possibly twisted DDPM denoising step. If not terminal denoising step loop back to computing fitness scores and repeat. By choosing a fitness score measuring whether samples are in distribution, the method favor more denoising trajectories with more precise denoising estimations hence more precise sampled actions.
• Figure 3: Each marker represents a unique combination of task, denoising-step count, and noise-injection scale, all at denoising step $t=5$. Clipping frequency is calculated as the proportion of entries in the flattened, noised action-sequence tensor that are projected onto the cube boundary. Grey lines show fourth-order polynomial regressions fitted separately for each task.
• Figure 4: Diffusion-generated face without noise injection.
• Figure 5: (a) Clip frequency normalized by the maximum observed value as a function of the number of denoising steps; clipping increases with more steps. (b) Impact of noise injection scale across action horizons, averaged over tasks; lower injection improves performance systematically. (c) Normalized score across horizons for tuned EDM in the few-step regime (for $n>5$, curves match $n=5$).