OMP: One-step Meanflow Policy with Directional Alignment

2025-12-22

AI Summary

To address high inference latency, architectural complexity, and poor few-shot generalization in generative policies for robotic dexterous manipulation, this paper proposes MeanFlow++, a mean-flow policy enhancement framework. Methodologically, it (1) introduces a cosine-direction alignment loss to decouple velocity direction and magnitude calibration, and (2) models dynamic trajectory evolution via differential derivative equations (DDEs), coupled with Jacobian-vector product (JVP)-based optimization to replace fixed-temperature denoising losses—enabling cooperative alignment between predicted mean velocities and ground-truth trajectories. The framework preserves single-step inference while significantly reducing computational overhead. Evaluated on Adroit and Meta-World benchmarks, MeanFlow++ achieves higher average success rates than MP1 and FlowPolicy, with particularly notable gains on challenging Meta-World tasks. It thus strikes an effective balance between real-time execution and trajectory fidelity.

Technology Category

• Intelligent Robots: Manipulation
• Search and Optimization: Sampling/Simulation-based Search
• Machine Learning: Imitation Learning & Inverse Reinforcement Learning

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications

Abstract

Robot manipulation, a key capability of embodied AI, has turned to data-driven generative policy frameworks, but mainstream approaches like Diffusion Models suffer from high inference latency and Flow-based Methods from increased architectural complexity. While simply applying meanFlow on robotic tasks achieves single-step inference and outperforms FlowPolicy, it lacks few-shot generalization due to fixed temperature hyperparameters in its Dispersive Loss and misaligned predicted-true mean velocities. To solve these issues, this study proposes an improved MeanFlow-based Policies: we introduce a lightweight Cosine Loss to align velocity directions and use the Differential Derivation Equation (DDE) to optimize the Jacobian-Vector Product (JVP) operator. Experiments on Adroit and Meta-World tasks show the proposed method outperforms MP1 and FlowPolicy in average success rate, especially in challenging Meta-World tasks, effectively enhancing few-shot generalization and trajectory accuracy of robot manipulation policies while maintaining real-time performance, offering a more robust solution for high-precision robotic manipulation.

Problem

• Improves few-shot generalization in robot manipulation policies
• Enhances trajectory accuracy while maintaining real-time performance
• Addresses misaligned velocity predictions and fixed temperature hyperparameters

Innovation

• Lightweight Cosine Loss aligns velocity directions
• DDE optimizes Jacobian-Vector Product operator
• Maintains real-time performance with enhanced generalization