PaperlayerView exact revision on arXiv

Source-linked AI summary

FlowPolicy: Enabling Fast and Robust 3D Flow-based Policy via Consistency Flow Matching for Robot Manipulation

Qinglun Zhang, Zhen Liu, Haoqiang Fan, Guanghui Liu, Bing Zeng, Shuaicheng Liu

arXiv:2412.04987v2cs.RO

TL;DR

Diffusion-based policies can model complex robotic trajectories, but repeated sampling creates an efficiency-quality trade-off, and conditional consistency flow matching with complex 3D vision remains underexplored. FlowPolicy conditions consistency flow matching on 3D point clouds to generate actions in one step. It reports a 7× reduction in average inference time while maintaining competitive success rates across Adroit and MetaWorld tasks.

  • Problem

    Diffusion policies model complex trajectories but require repeated inference sampling, and conditional consistency flow matching with complex 3D visual conditions remains underexplored.

    Source: Introduction · p. 2

  • Method

    FlowPolicy conditions a consistency flow model on 3D point clouds and robot-state embeddings, learning velocity-consistent straight-line flows for action generation.

    Source: FlowPolicy · p. 4, Consistency Flow Matching · p. 4

  • Results

    7× reduction in average inference time while maintaining competitive average success rates compared to state-of-the-art methods across 37 Adroit and MetaWorld tasks.

    Source: Introduction · p. 2, Conclusion · p. 7

  • Takeaways & Limitations

    FlowPolicy demonstrates real-time, one-step generation of high-quality robot actions with 3D conditioning across manipulation tasks.

    Source: Introduction · p. 2, Conclusion · p. 7

Abstract

from arXiv · show

Robots can acquire complex manipulation skills by learning policies from expert demonstrations, which is often known as vision-based imitation learning. Generating policies based on diffusion and flow matching models has been shown to be effective, particularly in robotic manipulation tasks. However, recursion-based approaches are inference inefficient in working from noise distributions to policy distributions, posing a challenging trade-off between efficiency and quality. This motivates us to propose FlowPolicy, a novel framework for fast policy generation based on consistency flow matching and 3D vision. Our approach refines the flow dynamics by normalizing the self-consistency of the velocity field, enabling the model to derive task execution policies in a single inference step. Specifically, FlowPolicy conditions on the observed 3D point cloud, where consistency flow matching directly defines straight-line flows from different time states to the same action space, while simultaneously constraining their velocity values, that is, we approximate the trajectories from noise to robot actions by normalizing the self-consistency of the velocity field within the action space, thus improving the inference efficiency. We validate the effectiveness of FlowPolicy in Adroit and Metaworld, demonstrating a 7$\times$ increase in inference speed while maintaining competitive average success rates compared to state-of-the-art methods. Code is available at https://github.com/zql-kk/FlowPolicy.

Introduction

Prior policy-learning methods face limits in modeling diverse actions or sampling efficiently, while conditional consistency flow matching with complex 3D vision remains underexplored. FlowPolicy applies 3D conditioning and consistency flow matching to generate actions in one step, reporting a 7× inference-time reduction with competitive success rates.

Introduction

  • Regression approaches struggle with multimodal behaviors and high-precision tasks, while energy-based models face training-stability challenges from negative-sample extraction.These limitations motivate generative policy approaches.Source: Introduction · p. 1
  • Diffusion policies model complex, high-dimensional trajectories, but their repeated inference sampling leaves efficiency and policy quality in tension.This runtime inefficiency limits practical application of learned policies.Source: Introduction · pp. 1–2
  • Conditional consistency flow matching for embodied tasks with complex 3D visual conditions remains underexplored.FlowPolicy addresses this gap with 3D visual conditioning.Source: Introduction · p. 2
  • FlowPolicy reduces average inference time by 7× while maintaining competitive average success rates against state-of-the-art 3D-based methods.The evaluation covers 37 Adroit and MetaWorld tasks.Source: Introduction · p. 2
  • FlowPolicy introduces conditional consistency flow matching for 3D manipulation, using straight-line flows and multi-segment training for one-step action generation.The framework conditions on 3D point clouds and targets a better efficiency-effectiveness trade-off.Source: Introduction · p. 2

Related Work

Prior diffusion and flow-matching work offers powerful generative approaches, but recursive sampling, distillation requirements, and limited evidence for complex 3D conditions constrain their relevance to robot manipulation. The paper situates FlowPolicy at the intersection of efficient consistency-based generation and 3D-conditioned policies.

Application of Diffusion Models to 3D Robotics

  • Diffusion models can require hundreds of inference steps, while consistency models support one-step generation but typically require distillation; FlowPolicy requires neither numerical simulation nor distillation.The cited discussion frames these requirements as constraints for real-world robotics.Source: Application of Diffusion Models to 3D… · p. 2

Flow Matching

  • Flow matching estimates a vector field defining an ODE and probabilistic path, but its use for complex robot actions with 3D spatial representations is rarely reported.Consistency-FM imposes velocity consistency on straight-line flows and uses multi-segment training to trade off sampling quality and efficiency.Source: Flow Matching · pp. 2–3

Method

FlowPolicy applies consistency flow matching to 3D visual imitation learning, aiming to generate high-quality robot actions in one step. Its method combines shared-endpoint straight-line flows, point-cloud conditioning, and multi-segment training.

Consistency Flow Matching

  • Consistency-FM defines straight-line flows from different times to a shared endpoint while constraining velocity to remain consistent along the flow.Its training loss also regularizes velocity and learns the data distribution; positive segment weights balance sections, especially harder middle sections.Source: Consistency Flow Matching · p. 4
  • Multi-segment Consistency-FM divides time into uniform subsets and relaxes velocity consistency to make data transmission more flexible and efficient.Source: Consistency Flow Matching · p. 4

FlowPolicy

  • FlowPolicy uses DP3 as its underlying architecture, combining 3D point-cloud observations with state embeddings to learn a consistency vector field for action sequences.The method describes one-step generation of high-quality actions from these conditions.Source: FlowPolicy · p. 4
  • The visual pipeline converts single-view depth images into point clouds, downsamples them with farthest point sampling, then encodes them with a lightweight MLP.Source: 3D visual conditional representation · p. 4
  • 19.9 ms average inference time per step is reported as 7× faster than DP3 and 3× faster than SimpleDP3 across 37 Adroit and MetaWorld tasks.The caption reports one-step inference and averages results across three random seeds.Source: Table 1 · p. 5
  • Two-segment training learns consistent vector fields that form a linear path from noisy trajectories to target action trajectories.At inference, sampled source noise is passed through the flow model to predict action trajectories.Source: Conditional Trajectory Generation with… · p. 5

Experiments

FlowPolicy was evaluated against 3D and 2D policy baselines on Adroit and MetaWorld tasks. It combined competitive success with faster inference, and its performance varied less with demonstration count than DP3’s.

Dataset and Implementation Details

  • The evaluation spans 37 tasks across Adroit dexterous-hand manipulation and MetaWorld tasks of varied difficulty, using heuristic-generated expert demonstrations.The baselines include DP3, Simple DP3, and 2D approaches; success rates, inference time, and number of function evaluations are assessed over repeated runs.Source: Simulation Benchmarks · p. 5, Baselines · p. 5, Evaluation Metrics · p. 5
  • Each task uses ten expert demonstrations, with point clouds downsampled to 512 points for Adroit or 1,024 for MetaWorld.The point clouds and robot state form the conditional input; images are cropped to 84×84 pixels.Source: Implementation Details · pp. 5–6

Comparison with State-of-the-art Methods

  • 19.9ms average inference time is more than 7× faster than DP3’s 145.7ms and over 3× faster than Simple DP3’s 63.0ms.Runtime comparisons cover three Adroit dexterity tasks and 34 MetaWorld tasks.Source: Quantitative Comparisons on Runtime · p. 6
  • FlowPolicy’s overall success rate is 1.3% higher than DP3 and 2.6% higher than Simple DP3 with the same expert demonstrations.The comparison covers 37 Adroit and MetaWorld tasks.Source: Quantitative Comparisons on Success Rate · p. 6
  • In Pen-turning, FlowPolicy more closely matches the target pen angle, while in Pick-Place it locates and carries the cube to the target after DP3 fails.The examples are qualitative comparisons on one Adroit and one MetaWorld task.Source: Figure 3 · p. 6, Qualitative Comparisons on… · p. 7
  • Across representative tasks, FlowPolicy maintains a stable Button Press Wall policy and reaches higher success rates than DP3 and Simple DP3 on Reach Wall, Hammer, and dexterity tasks.Most tasks converged and leveled off after 500–1,000 of the 3,000 training epochs.Source: Figure 4 · p. 6, Analysis of Learning Curves · p. 7

Ablation Studies on Expert Demonstrations

  • With more demonstrations, both methods generally improve, but FlowPolicy performs better on the reported tasks while DP3 can saturate or decline, as in Coffee Pull.On Pick-Place, success improves as demonstrations increase for both methods; FlowPolicy also outperforms DP3 with limited demonstrations on Reach Wall and Hammer.Source: Ablation Studies on Expert… · p. 7, Figure 5 · p. 7

Conclusion

The paper introduces FlowPolicy, a 3D point-cloud-conditioned visual imitation-learning method using consistency flow matching. Across 37 Adroit and MetaWorld tasks, it reports competitive success rates with real-time inference.

Conclusion

  • FlowPolicy is a 3D flow-based visual imitation-learning algorithm based on point clouds and consistency flow matching.The paper explores conditional generation with consistency flow matching under complex 3D visual representations.Source: Conclusion · p. 7
  • Straight-line flows accelerate transfer from noise to the action space, improving inference efficiency to real-time levels for robotic manipulation.Source: Conclusion · p. 7
  • Evaluation across 37 diverse Adroit and MetaWorld tasks demonstrates competitive success rates with real-time inference speed.Source: Conclusion · p. 7

Related papers

Loading 2412.04987v2…