Research map / Image generation / Fast diffusion sampling
Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Genetic Diffusion Policy (GDP), a population-based genetic denoising sampler with an out-of-distribution fitness score, plus schedule and noise-scale tweaks, enabling off-the-shelf diffusion policies to run with 2-5 denoising steps on Adroit and Robomimic tasks.
Problem
Diffusion policies for robotic manipulation require many sequential denoising steps, which is costly for real-time control; default inference also suffers from clipping-induced out-of-distribution intermediate states that degrade action quality at low step counts.
Approach
Reduce the injected noise scale and tailor the denoising schedule to low-dimensional action distributions, then apply a genetic (multinomial selection) population sampler whose fitness is an OoD score (Stein-type noise norm or clip-based), keeping in-distribution denoising trajectories without retraining the policy.
Key results
- On Adroit at 2 denoising steps, GDP (gamma=0.2) reaches 1.00 Hammer, 0.98 Relocate, 0.91 Pen, 1.00 Door, matching or exceeding 100-step DDPM and shortcut baselines.
- Shortcut models at 1 step reach only 0.83/0.93/0.74/0.89 on Adroit tasks, showing a performance gap relative to the proposed 2-step method.
- On Robomimic 2-step, GDP achieves 1.00 on Lift and 0.99 on Can PH, comparable to 100-step DDPM, though the paper notes most differences fall within 2 standard deviations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use 2-step GDP instead of DDPM + Schedule | DDPM policy · Adroit Hand tasks | normalized success rate: 0.74 -> 0.98 | Helped |
| Use 2-step GDP instead of the adapted DDPM schedule | DDPM policy · Robomimic | normalized success rate | No clear effect |
| Reduce noise scale to gamma 0 (ablation) | DDPM policy · Adroit Hand tasks | normalized success rate: 0.68 -> 0.99 | Helped |
| Use the adapted denoising schedule (ablation) | DDPM policy · Adroit Hand tasks, 2 steps | normalized success rate: 0.00 -> 0.87 | Helped |
| Add reduced noise scale gamma 0.2 (ablation) | DDPM + Schedule · Adroit Hand tasks, 2 steps | normalized success rate: 0.87 -> 0.98 | Helped |
| Reduce the number of denoising steps (ablation) | DDPM policy · Adroit Hand and Robomimic | normalized success rate | Mixed |
| Reduce noise injection (ablation) | DDPM policy · Adroit Hand and Robomimic | normalized success rate | Helped |
| Replace DDPM sampling with DDIM (ablation) | DDPM policy · Adroit Hand | normalized success rate: 0.69 -> 0.38 | Hurt |
| Use an EDM policy instead of a DDPM policy (ablation) | Diffusion policy · Adroit Hand, varying action horizons | success rate | Mixed |
| Use clipping statistics instead of the Stein score as the genetic selection heuristic (ablation) | GDP · Adroit Hand | normalized score | Hurt |
Surprising findings
- Reducing noise injection (gamma) improves performance in robotic tasks, contrary to findings in image generation.
- EDM policies were highly unreliable across action horizons, and DDIM performed poorly on Adroit despite being used for fast sampling in images.
Limitations
- Genetic algorithm is the simplest local version with no crossover or sophisticated mutation; memory and overhead grow with population size.
- Robomimic results show GDP mostly within 2 standard deviations of the adapted-schedule baseline, so gains over simple schedule tweaks are limited there.
- Theoretical analysis is preliminary and no formal theorems are proven; the method is argued to be specific to embodied AI rather than image generation.
Experimental setup
Datasets: Adroit Hand (D4RL) Robomimic (Lift, Can, Square, Transport, ToolHang) CelebA-HQ (ddpm-celebahq-256 for illustration)
Benchmarks: Adroit Hand (Hammer, Relocate, Pen, Door) Robomimic (Lift, Can, Square, Transport, ToolHang)
Metrics: normalized success rate clipping frequency normalized score
Baselines: DDPM DDIM EDM Shortcut models DDPM + Schedule
Builds on: Diffusion Policy DDIM EDM Shortcut models InferGrad
Closest papers by meaning
- Prediction with Action: Visual Policy Learning via Joint Denoising Process (NeurIPS 2024 · similarity 0.92)
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.92)
- Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.91)
- Efficient Diffusion Transformer Policies with Mixture of Expert Denoisers for Multitask Learning (ICLR 2025 · similarity 0.90)
- Diffusion-EDFs: Bi-equivariant Denoising Generative Modeling on SE(3) for Visual Robotic Manipulation (CVPR 2024 · similarity 0.90)
- ET-SEED: EFFICIENT TRAJECTORY-LEVEL SE(3) EQUIVARIANT DIFFUSION POLICY (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.