Research map / Image generation / Autoregressive and latent generation
READ: Retrieval-Enhanced Asymmetric Diffusion for Motion Planning
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 READ, a retrieval-enhanced diffusion framework for image-based robot motion planning that interpolates from a retrieved motion in a latent space (forward) and refines in the original task space (reverse via cold diffusion).
Problem
Image-based robot motion planning needs motions that are both kinematically feasible and task-successful; prior retrieval-based diffusion methods use long forward-reverse diffusion paths and struggle to refine retrieved motions accurately.
Approach
Directly diffuses from the retrieved motion to the target motion by interpolating in a low-dimensional latent space of a motion autoencoder (forward), while the reverse process runs in the original high-resolution task space with cold-diffusion-style one-step refinement, yielding asymmetric diffusion.
Key results
- READ achieves 70.2% Avg16 and 88.8% Avg12 success rate, outperforming R2-Diff (62.9 / 81.0) and RVT (65.9 / 74.3) on RLBench.
- Ablation: READ-L (latent-only diffusion) drops to 38.5 Avg16, and READ-O (original-space only) reaches 60.1 Avg16, showing the benefit of asymmetric diffusion.
- One-step refinement (N=1) already gives 68.8 Avg16, with iterative refinement reaching 70.2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use READ instead of the strongest listed baseline, RVT | RLBench manipulation planning · RLBench Avg16 | success rate: 65.9 -> 70.2 | Helped |
| use READ instead of the strongest listed baseline, R2-Diff | RLBench manipulation planning · RLBench Avg12 | success rate: 81.0 -> 88.8 | Helped |
| add asymmetric diffusion: latent-space forward process and original-space reverse process (ablation) | READ · RLBench Avg16 | success rate: 60.1 -> 70.2 | Helped |
| add asymmetric diffusion: latent-space forward process and original-space reverse process (ablation) | READ · RLBench Avg12 | success rate: 78.5 -> 88.8 | Helped |
| add motion refinement after retrieval (ablation) | READ · RLBench Avg16 | success rate: 56.6 -> 70.2 | Helped |
| add motion refinement after retrieval (ablation) | READ · RLBench Avg12 | success rate: 73.5 -> 88.8 | Helped |
| increase refinement iterations from 1 to 10 (ablation) | READ · RLBench Avg16 | success rate: 70.4 -> 71.0 | Helped |
| increase refinement iterations from 1 to 10 (ablation) | READ · RLBench Avg12 | success rate: 88.6 -> 89.8 | Helped |
| use cold diffusion instead of Euler Maruyama (ablation) | READ-O · RLBench Avg16 | success rate: 60.1 -> 58.8 | Hurt |
| use cold diffusion instead of Euler Maruyama (ablation) | READ-O · RLBench Avg12 | success rate: 78.5 -> 77.4 | Hurt |
| set k=10 rather than the default k=3 (ablation) | READ · RLBench Avg16 | success rate: 70.2 -> 67.1 | Hurt |
| set k=1 rather than the default k=3 (ablation) | READ · RLBench Avg16 | success rate: 70.2 -> 67.1 | Hurt |
Surprising findings
- Euler-Maruyama sampling slightly outperformed cold diffusion when used in the original space (60.1 vs 58.8 Avg16), suggesting cold diffusion does not improve refinement quality by itself.
Limitations
- Evaluated only on simulated RLBench manipulation tasks.
- Performance depends on retrieval quality and the retrieval database; a larger k can create train-inference distribution gaps.
- Iterative refinement gives only small gains over one-step refinement.
Experimental setup
Datasets: RLBench image-motion pairs (1000 training, 100 test)
Benchmarks: RLBench (16 manipulation tasks; Avg12 and Avg16 subsets)
Metrics: success rate
Baselines: Deterministic DDPM RVT VPSDE VINN DMO-EBM VPSDE+CG R2-Diff
Builds on: IR-SDE cold diffusion VPSDE R2-Diff VINN
Closest papers by meaning
- Diffusion Models for Robotics (AAAI 2025 · similarity 0.88)
- Model-based Diffusion for Trajectory Optimization (NeurIPS 2024 · similarity 0.86)
- Multi-Robot Motion Planning with Diffusion Models (ICLR 2025 · similarity 0.86)
- Generate Subgoal Images before Act: Unlocking the Chain-of-Thought Reasoning in Diffusion Model for Robot Manipulation with Multimodal Prompts (CVPR 2024 · similarity 0.86)
- Readout Guidance: Learning Control from Diffusion Features (CVPR 2024 · similarity 0.85)
- Motion Guidance: Diffusion-Based Image Editing with Differentiable Motion Estimators (ICLR 2024 · similarity 0.85)
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.