Date of Award
2026
Document Type
Open Access Dissertation
Degree Name
Doctor of Philosophy in Mechanical Engineering-Engineering Mechanics (PhD)
Administrative Home Department
Department of Mechanical and Aerospace Engineering
Advisor 1
Jung Yun Bae
Committee Member 1
Darrell L. Robinette
Committee Member 2
Tan Chen
Committee Member 3
Myoungkuk Park
Abstract
Multi-robot systems (MRS) are playing an increasingly significant role in dynamic and highly constrained environments, ranging from time-critical Urban Search and Rescue (USAR) operations to high-density automated warehouses. Effective Multi-Robot Task Allocation (MRTA) and path planning are essential to the operational success of these systems. However, traditional exact mathematical solvers scale poorly as fleet sizes grow, and rule-based heuristics often fail to adapt to complex heterogeneous constraints or prevent severe traffic congestion. While recent advancements in Large Language Models (LLMs) and Deep Reinforcement Learning (DRL) offer unprecedented adaptive reasoning, their direct application to physical robotic systems is hindered by inherent stochasticity, a reliance on manual prompt engineering, and extreme sample inefficiency when processing complex spatial topologies.
This dissertation proposes a unified, multi-layered framework that bridges high-level artificial intelligence with strict physical constraints to achieve scalable, reliable, and deterministic multi-robot coordination. First, it investigates the zero-shot reasoning capabilities of LLMs, demonstrating that structured prompt engineering can successfully allocate tasks and plan paths for heterogeneous robots in dynamic USAR environments. To overcome the unreliability and rigidity of static manual prompting, a novel Reinforcement Learning (RL) architecture is then introduced to autonomously optimize LLM inputs. By training an agent to dynamically adjust structural and semantic instructions based on continuous simulation feedback, the system successfully mitigates LLM stochasticity, guarantees the synthesis of executable control code, and significantly outperforms traditional exact solvers in both scalability and secondary constraint adherence.
Finally, to address the challenge of spatial reasoning in dynamic environments with mazes, this work introduces a Self-Supervised Learning (SSL) framework that decouples topological understanding from combinatorial dispatching. By pre-training a deep encoder via self supervised learning to understand the physical environment and subsequently fine-tuning a centralized RL policy, the resulting End-to-End model naturally mitigates traffic congestion without requiring computationally heavy hierarchical architectures. Ultimately, this dissertation establishes a robust, non-hierarchical paradigm that synergizes natural language reasoning, reinforcement learning, and spatial representation to enable fully autonomous, highly efficient multi-robot operations in complex environments.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Kannan, Kaushik, "Heterogeneous Multi-Robot Coordination In Dynamic Environments", Open Access Dissertation, Michigan Technological University, 2026.
Included in
Controls and Control Theory Commons, Navigation, Guidance, Control, and Dynamics Commons, Robotics Commons