Distributed Koopman Operator Learning for Perception and Safe Navigation

TL;DR

This paper introduces a scalable, distributed Koopman operator learning framework integrated with model predictive control for safe, efficient autonomous navigation in dynamic environments, leveraging multi-agent perception and obstacle prediction.

Contribution

It develops a novel distributed Koopman learning algorithm that enables multi-agent obstacle prediction and safe navigation without centralized data, with theoretical guarantees and practical validation.

Findings

Supports large-scale multi-agent obstacle prediction

Ensures collision-free navigation through convex safety constraints

Demonstrates reliable performance in complex simulations

Abstract

This paper presents a unified and scalable framework for predictive and safe autonomous navigation in dynamic transportation environments by integrating model predictive control (MPC) with distributed Koopman operator learning. High-dimensional sensory data are employed to model and forecast the motion of surrounding dynamic obstacles. A consensus-based distributed Koopman learning algorithm enables multiple computational agents or sensing units to collaboratively estimate the Koopman operator without centralized data aggregation, thereby supporting large-scale and communication-efficient learning across a networked system. The learned operator predicts future spatial densities of obstacles, which are subsequently represented through Gaussian mixture models. Their confidence ellipses are approximated by convex polytopes and embedded as linear constraints in the MPC formulation to guarantee safe and collision-free navigation. The proposed approach not only ensures obstacle avoidance but also scales efficiently with the number of sensing or computational nodes, aligning with cooperative perception principles in autonomous navigation applications. Theoretical convergence guarantees and predictive constraint formulations are established, and extensive simulations demonstrate reliable, safe, and computationally efficient navigation performance in complex environments.