Integrating Flow Matching and Model Predictive Control for Safe Robot Navigation

I. INTRODUCTION

• Generative paths can become unsafe when new obstacles appear or dynamics and constraints change, motivating a lower-level controller that adapts to current information.The cited example includes changes to velocity limits.Source: §I · p. 1
• MPPFC tracks and locally reshapes generated paths, enforcing state and input constraints while allowing temporary deviations from the reference.Unlike tracking MPC, it optimizes path progression and velocity for adaptive motion profiles and reactive behavior.Source: §I · p. 1
• Unlike related diffusion-and-MPC approaches described here, this method uses MPPFC as the main controller rather than a filter that modifies unsafe inputs.The authors say the underlying MPC methods in those works do not guarantee safety or path-following.Source: §I · p. 1
• The framework combines a flow-matching planner learned from human demonstrations with MPPFC to enforce constraints and avoid collisions during execution.Predictive-control solutions also bias the planner during event-triggered replanning.Source: §I · p. 1
• The evaluation examines collision-free path-following in obstacle configurations not observed during training.The paper also frames its study around navigation with unknown obstacles and the effects of horizon expansion and replanning.Source: §I · p. 2

II. PROBLEM FORMULATION

• The objective is to navigate a cluttered environment while satisfying system constraints and avoiding collisions, using expert demonstrations to encode spatial-navigation preferences.The formulation learns a flow-matching planner rather than relying on classical sampling-based planners such as RRT.Source: §II · p. 2
• The planner samples candidate paths from a learned flow-matching model conditioned on start and goal configurations.Paths are geometric curves in output space represented as linear combinations of fixed-order B-splines, with the model trained on their coefficients.Source: §II · p. 2
• Because the generative planner does not enforce real-time constraints or active collision avoidance from all current local information, that responsibility goes to a lower-level controller.The generated path remains a reference prior for this controller.Source: §II · p. 2
• Parallel MPPFC instances with different horizons adapt paths online to environmental changes and unseen obstacles, while predefined criteria trigger resampling biased by the controller’s solution trajectory.This coupling is intended to keep global planning consistent with local execution and reduce corrective effort.Source: §II · p. 2

III. GENERATIVE MODELING FOR PATH PLANNING

• The section introduces path parametrization, demonstration conversion, and a probabilistic generative model over paths inferred from expert behavior.The model is described as learning path distributions from demonstrations.Source: §III · p. 2

A. From Trajectory Data to Path Parametrization

• Each output path is represented as a linear combination of M clamped B-splines of order r, with Mny governing coefficients specifying the full path.The basis functions are defined by the Cox-de Boor recursion formula.Source: A. From Trajectory Data to Path… · p. 2
• Time-indexed expert demonstrations are converted into time-independent geometric paths and fitted with least-squares regression to obtain coefficient data.The representation is agnostic to trajectory length, sampling rate, and velocity profile; evaluation points are normalized by demonstration duration.Source: A. From Trajectory Data to Path… · p. 2

B. Modeling and Learning Expert Behaviors

• The shortcut model [19] learns the distribution of path coefficients conditioned on context representing start and goal states, capturing potentially multimodal expert behavior.Its time-varying vector field maps samples from initial noise to the data distribution.Source: B. Modeling and Learning Expert… · p. 2
• Training matches displacement between noise and demonstration samples while enforcing consistency across integration steps.The resulting model can efficiently generate expert-like path coefficients.Source: B. Modeling and Learning Expert… · p. 2
• The pipeline samples path parametrizations, selects a collision-free path for known obstacles, and uses parallel MPPFC instances with varying horizons to follow it.If an unforeseen obstacle makes the path or control problem infeasible, the shortest feasible horizon is applied before replanning under the stated condition.Source: Figure 1 · p. 3

A. Model Predictive Path-Following Control

• MPPFC steers the system along a given path using parallel controllers with incrementally extended prediction horizons.Each controller's horizon is T + iδt, where T is the base horizon and δt is a fixed increment.Source: A. Model Predictive Path-Following… · p. 3
• State and input constraints capture physical limits and collision avoidance, while virtual dynamics and input control progress along the reference path.The terminal cost promotes consistency between true and virtual states at the horizon end, supporting convergence while respecting system and path constraints.Source: A. Model Predictive Path-Following… · p. 3
• The stage cost penalizes path deviation and large control inputs while encouraging forward progress toward the path's final parameter.The terminal cost and constraint jointly ensure convergence to the path when appropriately chosen; optimization can use gradient-based or sampling-based methods, though sampling-based methods do not guarantee strict constraint satisfaction.Source: A. Model Predictive Path-Following… · p. 3

B. MPPFC for Safety Filtering

• Directly guiding the flow model is limited by costly evaluation: too few integration steps make strong guidance risk leaving the data distribution and weak guidance have little effect.Updating the global plan for purely local changes is also computationally inefficient, motivating local MPPFC re-planning.Source: B. MPPFC for Safety Filtering · p. 4
• Runtime sensor detections are incorporated as state constraints in the safe set, then removed after the robot passes the obstacle and it leaves sensing range.The framework assumes sensors such as LiDAR and convexifies newly detected obstacles.Source: B. MPPFC for Safety Filtering · p. 4
• When an obstacle forces a large deviation, parallel MPPFC instances extend the horizon until a feasible solution is found, executing the shortest feasible horizon.The controller returns to the shortest horizon once local feasibility is reestablished; fixed collocation points make the shorter horizon's integration step smaller and control fidelity higher.Source: B. MPPFC for Safety Filtering · p. 4
• Event-triggered re-planning begins after local safety is secured and the system exits a user-defined neighborhood around an evaded obstacle, restoring alignment with demonstrated behavior.The reference path can be switched without modifying the underlying algorithmic procedure.Source: B. MPPFC for Safety Filtering · pp. 4–5
• Moving obstacles require additional models to predict their behavior, unlike the obstacle handling described for newly detected static objects.The cited passage points to MPC strategies for such scenarios in autonomous driving.Source: B. MPPFC for Safety Filtering · p. 4
• The strategy prioritizes real-time adaptability and empirical performance over conservative formal constraints, assuming a topologically viable alternative path exists.The paper describes collision-free behavior in practice under this assumption.Source: B. MPPFC for Safety Filtering · p. 5

C. Biasing the Planner Toward the MPPFC Solution

• The MPC prediction is fit with a B-spline to approximate a clean path, then used to initialize flow matching at an intermediate noise level.The true clean sample is generally unavailable; the resulting candidate paths are locally aligned with the controller's current plan and likely maintain its collision-avoidance guarantees locally.Source: C. Biasing the Planner Toward the… · p. 5
• Biasing toward the MPC prediction typically requires fewer refinement steps and can reduce corrective actions and MPC solving effort.The bias increases the likelihood of obtaining a collision-free alternative reference path.Source: C. Biasing the Planner Toward the… · p. 5

V. SIMULATION EXPERIMENTS

• The evaluation is conducted in simulated tasks.Source: §V · p. 5
• One experiment question concerns system behavior during navigation with unknown obstacles.Source: §V · p. 5
• The experiments also examine how MPC horizon expansion and re-planning strategies influence success rate.Source: §V · p. 5

A. Environment, Data Collection, and Shortcut Model

• Approximately 1,800 paths were collected as human demonstrations in MuJoCo, where random start-goal pairs produced multimodal trajectories.The shortcut model uses a diffusion-transformer architecture.Source: A. Environment, Data Collection, and… · p. 5

B. Dynamics, Environments, and MPPFC Design

• The simulated robot has bounded translational and angular velocities, with a 50 ms sampling time.Its state comprises planar position and heading, and its output is planar position.Source: B. Dynamics, Environments, and MPPFC… · p. 5
• The robot moves among four obstacle configurations with identical known static obstacles and varying previously unknown obstacles.The MPPFC uses obstacles within sensing range to define a safe state-constraint set.Source: B. Dynamics, Environments, and MPPFC… · p. 5
• Fig. 4 depicts candidate paths, the selected reference, closed-loop execution, predicted motion, and re-planning paths around newly detected obstacles.Known obstacles are blue; previously unknown obstacles are red, with a red boundary marking the robot-size margin.Source: Figure 4 · p. 6
• Gradient-based MPPFC implementations were used because sampling-based inputs lacked smoothness and reduced path-following accuracy and traversal performance.The re-planning conditions are periodic resampling, predicted collision with a newly detected obstacle, or exiting the safety margin around an evaded obstacle.Source: B. Dynamics, Environments, and MPPFC… · p. 6

• The robot samples candidate paths from the shortcut model for a given start and goal, then executes the shortest-horizon feasible controller’s inputs for a randomly selected path.The sampled paths are collision-free relative to obstacles encountered during demonstrations.Source: C. Closed-Loop Performance · p. 6
• A newly sensed obstacle can make the reference path collide, so collision avoidance is added as a hard MPC constraint; terminal path realignment can then make every controller instance infeasible.Obstacle detection is simulated with LiDAR.Source: C. Closed-Loop Performance · p. 6
• When a new obstacle is constrained, the path parameter is initialized at the reference endpoint, and controller horizons can be extended iteratively until feasibility returns.The endpoint initialization biases solutions toward path progress and can increase the likelihood that a controller remains feasible while returning to the reference.Source: C. Closed-Loop Performance · pp. 6–7
• After the robot exits a user-defined neighborhood around an evaded obstacle, re-planning samples paths from its current position to the goal using noisy MPPFC output as the planner initialization.The biased planner produces locally collision-free candidates in two integration steps.Source: C. Closed-Loop Performance · p. 7
• The selected new path must avoid both currently sensed obstacles and demonstration obstacles; if none qualifies, parallel MPC instances continue handling local feasibility.The process repeats when further previously unknown obstacles appear, until the robot reaches its destination.Source: C. Closed-Loop Performance · p. 7

D. Validation using Ablation Studies

• The ablation tests K = {5, 7, 15} controller instances, corresponding to maximum horizons of 1.8 s, 2.2 s, and 3.8 s, with 10 simulations per scenario.Fig. 6 reports generative-planner re-planning events for successful trials.Source: D. Validation using Ablation Studies · p. 7, Figure 6 · p. 7
• Fig. 5 shows P3 achieves 100% success across configurations at Tmax = 2.2 s and 3.8 s; at 3.8 s, P2 reaches 90% in C4.At Tmax = 1.8 s, the controller does not always find a feasible solution; P3 also requires fewer generative-planner evaluations without compromising performance.Source: Figure 5 · p. 7, D. Validation using Ablation Studies · p. 7

VI. CONCLUSIONS

• The conclusion proposes adding richer perception context to path generation and extending the framework to dynamic environments through motion prediction.The reported framework combines expert-demonstration flow-matching planning with MPPFC for adaptive, collision-free execution around previously unseen obstacles.Source: §VI · p. 7