Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unified Generation-Refinement Planning: Bridging Guided Flow Matching and Sampling-Based MPC for Social Navigation

Kazuki Mizuta, Karen Leung

TL;DR

The paper tackles the problem of safe, real-time robot planning in dynamic human environments by bridging learning-based trajectory generation with optimization-based constraint enforcement. It introduces a reward-guided conditional flow matching (CFM) model to produce multimodal trajectory priors and couples it with model predictive path integral (MPPI) control, forming a bidirectional loop where CFM priors guide MPPI refinement and MPPI solutions warm-start subsequent CFM generation. Key contributions include the integration of CFM with MPPI, reward-guided, retraining-free adaptation of CFM, mode-selective MPPI to preserve multimodal options, and empirical validation in social navigation showing improved safety, goal attainment, and real-time performance. The approach offers a practical path to robust social navigation by combining expressive generative priors with explicit constraint handling, enabling real-time deployment in crowded environments.

Abstract

Planning safe and effective robot behavior in dynamic, human-centric environments remains a core challenge due to the need to handle multimodal uncertainty, adapt in real-time, and ensure safety. Optimization-based planners offer explicit constraint handling but performance relies on initialization 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.

Unified Generation-Refinement Planning: Bridging Guided Flow Matching and Sampling-Based MPC for Social Navigation

TL;DR

The paper tackles the problem of safe, real-time robot planning in dynamic human environments by bridging learning-based trajectory generation with optimization-based constraint enforcement. It introduces a reward-guided conditional flow matching (CFM) model to produce multimodal trajectory priors and couples it with model predictive path integral (MPPI) control, forming a bidirectional loop where CFM priors guide MPPI refinement and MPPI solutions warm-start subsequent CFM generation. Key contributions include the integration of CFM with MPPI, reward-guided, retraining-free adaptation of CFM, mode-selective MPPI to preserve multimodal options, and empirical validation in social navigation showing improved safety, goal attainment, and real-time performance. The approach offers a practical path to robust social navigation by combining expressive generative priors with explicit constraint handling, enabling real-time deployment in crowded environments.

Abstract

Planning safe and effective robot behavior in dynamic, human-centric environments remains a core challenge due to the need to handle multimodal uncertainty, adapt in real-time, and ensure safety. Optimization-based planners offer explicit constraint handling but performance relies on initialization 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.

Paper Structure

This paper contains 17 sections, 1 theorem, 11 equations, 14 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries: Conditional Flow Matching
  4. Problem Formulation
  5. Guided Conditional Flow Matching for Trajectory Optimization
  6. CFM model training
  7. CFM guidance with reward functions
  8. Reward function
  9. Bridging Flow Matching with Trajectory Optimization
  10. Benefits of integrating CFM with MPPI
  11. Mode selection: Handling multimodal priors from CFM
  12. Warmstarting CFM with MPC solutions
  13. Experiments and Discussion
  14. Experimental setup
  15. Qualitative results
  16. ...and 2 more sections

Key Result

Proposition 1

Given a base velocity field $v_\theta(\mathbf{z}_\tau,\mathbf{c},\tau)$ and a differentiable reward function $R(\mathbf{z}_1)$, the guided velocity field $\tilde{v}_\theta(\mathbf{z}_\tau,\mathbf{c},\tau)$ is given by where $\lambda_\text{guide}$ is a hyperparameter controlling guidance strength.

Figures (14)

  • Figure 1: Overview of the proposed unified planning framework for dynamic environments: At each planning step, our conditional flow matching (CFM) model generates context-aware and multimodal trajectory candidates guided by a reward function. Promising candidates are selected, then refined, and the best trajectory is selected and executed. The optimal trajectory warm-starts the next CFM generation for the next planning step.
  • Figure 2: A safety-guided conditional flow matching (CFM) model generates diverse trajectories as priors for model predictive control (MPC), which in turn warm-starts the next CFM sampling step.
  • Figure 3: Gaussian.
  • Figure 4: Diffusion.
  • Figure 5: CFM.
  • ...and 9 more figures

Theorems & Definitions (2)

  • Proposition 1: Reward-guided CFM Velocity Field
  • proof