Date of Graduation
Summer 2026
Degree
Master of Science in Computer Science
Department
Computer Science
Committee Chair
Siming Liu
Abstract
Enabling heterogeneous robots with diverse capabilities, roles, and task responsibilities to coordinate effectively in complex, dynamic environments remains a fundamental challenge in autonomous multi-robot systems. Multi-Agent Reinforcement Learning (MARL) provides a promising framework for decentralized cooperation. However, most existing MARL approaches assume homogeneous agents or fixed single-task settings and suffer significant performance degradation as the number of robot and task types increases. This thesis presents a scalable shared-policy MARL framework that allows heterogeneous robots to learn specialized behaviors for individual, sequential, and collaborative tasks involving temporal and spatial dependencies through a single neural policy. I first embed robot identity directly in the observation space through Signed Type Encoding and transfer learned individual skills to multi-step and multi-robot tasks via curriculum learning. This achieves stable specialization for small numbers of robot types, but does not scale beyond a few types and assumes fixed-order sequences and rigid one-robot-per-type collaboration. To overcome these limitations, I introduce a priority-aware grid sensing mechanism that embeds task relevance directly into perception: each grid cell is augmented with a distance-weighted signal indicating whether a detected task is relevant to the observing robot, allowing a shared policy to scale to as many as ten robot and task types. Building on this representation, I generalize sequential tasks to arbitrary length and ordering, and collaborative tasks to dynamic requirement vectors that permit any arrival order and multiple robots of the same type. Experiments in the Unity platform up to ten-type configurations demonstrate consistent learning, substantial scalability gains over a default grid-sensor baseline, and strong generalization to previously unseen task sequences and collaboration requirements without retraining. These results demonstrate that relevance-aware observation design significantly improves the scalability of heterogeneous MARL, providing a flexible and generalizable solution for real-world multi-robot coordination.
Keywords
multi-agent reinforcement learning, heterogeneous multi-robot systems, shared-policy learning, priority-aware grid sensing, curriculum learning, sequential and collaborative task coordination, transfer learning
Subject Categories
Robotics
Copyright
© Rehab Uddin Shawon
Recommended Citation
Shawon, Rehab Uddin, "Heterogeneous Robots Cooperation via Multi-Agent Reinforcement Learning" (2026). Graduate Theses/Dissertations. 4196.
https://bearworks.missouristate.edu/theses/4196
