Research map / Reinforcement learning / Visual and representation RL
Compose Your Policies! Improving Diffusion-based or Flow-based Robot Policies via Test-time Distribution-level Composition
Jiahang Cao, Yize Huang, Hanzhong Guo, Qiang Zhang, Rui Zhang, Weijian Mai, Mu Nan, Jiaxu Wang, Hao Cheng, Jingkai SUN, Gang Han, Wen Zhao and 6 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| compose DP with MP by convex score combination | Diffusion Policy · Robomimic+PushT average | success rate (%): 39.19 -> 41.41 | Improved |
| compose Florence-Policy-D with DP | Florence Policy-D · Robomimic+PushT average | success rate (%): 61.25 -> 66.76 | Improved |
| compose pi0 with Flow Policy | pi0 · Robomimic+PushT average | success rate (%): 86.42 -> 88.94 | Improved |
| compose Florence-Policy-F with Flow Policy | Florence Policy-F · Robomimic+PushT average | success rate (%): 78.84 -> 86.39 | Improved |
| compose RDT with DP point cloud | DP (point cloud) · RoboTwin 2.0 average | success rate: 0.65 -> 0.72 | Improved |
| compose DP image with DP point cloud | DP (point cloud) · RoboTwin 2.0 average | success rate: 0.65 -> 0.70 | Improved |
| tune convex weight when both policies are moderately accurate (ablation) | GPC DP img + DP pcd · Empty Cup Place | success rate: 0.62 -> 0.86 | Improved |
| compose with a much weaker policy (ablation) | GPC DP img + DP pcd · Pick Apple Messy | success rate: 0.26 -> 0.25 | No clear effect |
| compose with a much weaker policy using equal weights (ablation) | GPC DP img + DP pcd · Block Hammer Beat | success rate: 0.76 -> 0.12 | Hurt |
| give stronger policy larger weight (ablation) | GPC DP img + DP pcd · Dual Bottles Pick (Easy) | success rate: 0.77 -> 0.85 | Improved |
| add a third base policy (FP+FP-F+pi0) (ablation) | GPC · Robomimic Can | success rate: 0.99 -> 1.00 | Improved |
| compose DP and MP with logical AND superposition (ablation) | GPC · Robomimic+PushT average | success rate (%): +25.73%; SR 64.92 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Flow Policy — FlowPolicy: Enabling Fast and Robust 3D Flow-Based Policy via Consistency Flow Matching for Robot Manipulation (AAAI 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Diffusion Policy Policy Optimization (ICLR 2025 · similarity 0.90)
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.89)
- GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- DiffusionVLA: Scaling Robot Foundation Models via Unified Diffusion and Autoregression (ICML 2025 · similarity 0.89)
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.89)
- Conditioning Matters: Training Diffusion Policies is Faster Than You Think (NeurIPS 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.
