Research map / Reinforcement learning / Visual and representation RL
FlowPolicy: Enabling Fast and Robust 3D Flow-Based Policy via Consistency Flow Matching for Robot Manipulation
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 FlowPolicy, a 3D point-cloud-conditioned consistency flow matching policy that generates robot actions in one inference step for imitation learning on Adroit and Metaworld.
Problem
Diffusion and flow-matching policies for robot manipulation require many sampling steps, creating a trade-off between inference speed and action quality, especially with 3D visual conditioning.
Approach
Conditions a segmented consistency flow matching model (Consistency-FM) on 3D point cloud and state embeddings so that straight-line velocity-consistent flows from noise to actions allow single-step policy generation without distillation, built on DP3's architecture.
Key results
- FlowPolicy achieves 70.0% average success rate with one NFE on 37 Adroit and Metaworld tasks, versus 68.7% for DP3 (10 NFE).
- Average inference time of 19.9 ms per step versus 145.7 ms for DP3, roughly 7x faster.
- Compared with Simple DP3 (67.4% success, 63.0 ms), FlowPolicy is 2.6 points higher in success and about 3x faster.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use one-step consistency flow matching instead of DP3's iterative inference | FlowPolicy · Adroit + Metaworld (37 tasks), average inference time | inference time (ms): 145.7±2.3 -> 19.9 ± 0.2 | Helped |
| Use one-step consistency flow matching instead of Simple DP3's iterative inference | FlowPolicy · Adroit + Metaworld (37 tasks), average inference time | inference time (ms): 63.0±5.9 -> 19.9 ± 0.2 | Helped |
| Use consistency flow matching with 3D point-cloud conditioning instead of DP3 | FlowPolicy · Adroit + Metaworld (37 tasks), average success rate | success rate (%): 68.7±4.7 -> 70.0 ± 4.7 | Helped |
| Use consistency flow matching with 3D point-cloud conditioning instead of Simple DP3 | FlowPolicy · Adroit + Metaworld (37 tasks), average success rate | success rate (%): 67.4±5.0 -> 70.0 ± 4.7 | Helped |
| Increase the number of expert demonstrations (ablation) | FlowPolicy and DP3 · Adroit and Metaworld representative tasks | success rate | Helped |
| Increase the number of demonstrations on Coffee Pull (ablation) | DP3 · Coffee Pull | success rate | Hurt |
| Use FlowPolicy instead of DP3 with limited demonstrations (ablation) | FlowPolicy · Reach Wall and Hammer | success rate | Helped |
Surprising findings
- Consistency Policy and other 2D baselines score lower than 3D baselines, while the 2D diffusion policy is notably weaker than DP3 in the reported results.
Limitations
- Evaluated only in simulation on Adroit and Metaworld with heuristic expert demonstrations
- Relies on a limited set of 10 demonstrations per task
- Uses a two-segment Consistency-FM design choice without broader ablation of segment count
Experimental setup
Datasets: Adroit Metaworld
Benchmarks: Adroit Metaworld
Metrics: success rate inference time per step (ms) NFE
Baselines: Diffusion Policy DP3 Simple DP3 AdaFlow Consistency Policy
Builds on: DP3 Consistency-FM flow matching consistency models Diffusion Policy
Closest papers by meaning
- Frequency‑Aware Flow Matching for Continuous and Consistent Robotic Action Generation (NeurIPS 2026 · similarity 0.91)
- FlowRAM: Grounding Flow Matching Policy with Region-Aware Mamba Framework for Robotic Manipulation (CVPR 2025 · similarity 0.90)
- FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · similarity 0.90)
- Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.90)
- ReinFlow: Fine-tuning Flow Matching Policy with Online Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.88)
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.