Unified Generation-Refinement Planning: Bridging Flow Matching and Sampling-Based MPC

2025-08-02

AI Summary

Real-time robot motion planning in dynamic human environments faces the challenge of simultaneously modeling uncertainty, strictly satisfying safety and kinematic constraints, and generating diverse, context-appropriate behaviors. Method: This paper proposes a unified generative-optimization co-design framework. Its core innovation is a bidirectional feedback loop between a reward-guided Conditional Flow Matching (CFM) model and a Model Predictive Path Integral (MPPI) controller: CFM generates diverse, prior-informed initial trajectory proposals, while MPPI performs physics-based, constraint-aware trajectory optimization. Crucially, the two components interact via differentiable closed-loop integration, enabling synergistic trade-offs between generative diversity and optimization-driven safety. Results: Evaluated on autonomous social navigation tasks, the method achieves significant improvements in planning latency, safety compliance, and robustness to environmental dynamics—demonstrating effectiveness in complex, real-world human-centric scenarios.

Technology Category

• Humans and AI: Human-Aware Planning and Behavior Prediction
• Intelligent Robots: Motion and Path Planning
• Planning, Routing, and Scheduling: Replanning and Plan Repair

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Planning safe and effective robot behavior in dynamic, human-centric environments remains a core challenge due to the need to handle uncertainty, adapt in real-time, and ensure safety. Optimization-based planners offer explicit constraint handling but rely on oversimplified initialization, reducing solution quality. Learning-based planners better capture multimodal possible solutions but struggle to enforce constraints such as safety. In this paper, we introduce a unified generation-refinement framework bridging learning and optimization with a novel reward-guided conditional flow matching (CFM) model and model predictive path integral (MPPI) control. Our key innovation is in the incorporation of a bidirectional information exchange: samples from a reward-guided CFM model provide informed priors for MPPI refinement, while the optimal trajectory from MPPI warm-starts the next CFM generation. Using autonomous social navigation as a motivating application, we demonstrate that our approach can flexibly adapt to dynamic environments to satisfy safety requirements in real-time.

Problem

• Planning safe robot behavior in dynamic human environments
• Balancing learning-based and optimization-based planning methods
• Ensuring real-time adaptation and safety in robot navigation

Innovation

• Unified generation-refinement framework combining learning and optimization
• Reward-guided conditional flow matching for informed priors
• Bidirectional exchange between CFM and MPPI for refinement