Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Maryam Khazaei

Second Advisor

Navrati Saxena

Third Advisor

Nishchay Nagendra

Keywords

Multi-agent Path Optimization, Q-learning, Visibility Graph, Convex Optimization, Piecewise Quadratic Bézier Path, Adaptive Path Planning

Abstract

Robot navigation in a multi-agent setting requires a balance between safety and efficiency, especially in dense environments. In these two-dimensional spaces, the scope for geometric errors is much less and could lead to collisions or immobility. This project proposes to address the navigation task using a two-phase path planning pipeline that combines reinforcement learning and convex optimization in a scalable and robust manner. The first phase consists of generating diverse collision-free paths using a Q-learning agent that is trained on a visibility graph representation of the environment. The discretization of the environment using waypoint-based graphs allows the agent to train and learn without requiring manual heuristics, as the structural integrity of the scenario is maintained. This discretization process allows the agent to generate multiple collision-free paths while remaining robust and scalable across different start- goal and obstacle variations. The resultant paths from the first phase are piecewise linear and discrete, but they avoid obstacles and ensure path diversity, which are essential for multi-agent systems. In the second phase, the paths produced by the Q-learning agent are refined using Multi-Agent Interleaving Convex Optimization(MICO). This method transforms each discrete path into a continuous representation using quadratic Bézier curves that are optimized to reduce curvature and path length while maximizing obstacle and agent clearance. This optimization strategy iteratively selects the paths to be refined using priority scores while ensuring path diversity is maintained. This results in smooth and geometrically consistent trajectories that are suitable for actuation by real robots. Overall, this framework demonstrates that using structure-aware Q-learning on visibility graphs with convex Bézier-based optimization yields effective results that convert adaptive exploration to execution-ready paths. The hybrid approach shows how learning-based diversity and refinement-based precision can be combined into a scalable and robust multi-agent path planning pipeline that ensures smoothness and safety.

Download

Share

COinS