Distributed coordination and perception for multi-robot systems
File(s)
Author(s)
Patwardhan, Aalok
Type
Thesis
Abstract
This thesis explores a scalable, fully distributed framework for multi-robot collaboration under uncertainty in real-world environments. Multi-robot systems offer significant advantages over single-robot solutions, including parallelised task execution, redundancy, and scalability, enabling applications such as autonomous exploration and environmental monitoring.
Achieving these benefits requires coordination, yet centralised approaches face bottlenecks, single points of failure, and limited adaptability. Real-world tasks are further complicated by noisy sensors, partial observations, and unreliable communication. Distributed frameworks overcome these challenges by enabling robots to reason locally and communicate with neighbours, supporting scalable and adaptive collective behaviour.
We formulate multi-robot coordination as probabilistic inference on factor graphs, using Gaussian Belief Propagation to perform local, asynchronous message-passing over nodes representing robot states and constraints. Our framework addresses complex multi-robot problems spanning multiple competencies such as path planning, exploration, and shape formation tasks. Robots coordinate to optimise locally, while emergent global behaviour arises from sparse but structured interactions. The framework also supports distributed consensus over continuous and discrete decision spaces, allowing swarms to agree on shared global states. Practical applicability is demonstrated via deployment on low-cost embedded hardware, achieving fully decentralised coordination through peer-to-peer communication. Finally, the thesis also demonstrates the applications of factor graphs to computer vision tasks, developing an uncertainty-aware rotation estimator that produces temporally consistent camera rotation estimates from RGB images.
This thesis demonstrates that factor-graph-based methods provide a lightweight and generalisable backbone for multi-robot systems, by unifying planning, exploration, formation, and consensus under a distributed, probabilistic approach. We envisage that these methods unlock the potential for large, heterogeneous swarms of the future to perform complex, coordinated behaviours in dynamic and uncertain environments.
Achieving these benefits requires coordination, yet centralised approaches face bottlenecks, single points of failure, and limited adaptability. Real-world tasks are further complicated by noisy sensors, partial observations, and unreliable communication. Distributed frameworks overcome these challenges by enabling robots to reason locally and communicate with neighbours, supporting scalable and adaptive collective behaviour.
We formulate multi-robot coordination as probabilistic inference on factor graphs, using Gaussian Belief Propagation to perform local, asynchronous message-passing over nodes representing robot states and constraints. Our framework addresses complex multi-robot problems spanning multiple competencies such as path planning, exploration, and shape formation tasks. Robots coordinate to optimise locally, while emergent global behaviour arises from sparse but structured interactions. The framework also supports distributed consensus over continuous and discrete decision spaces, allowing swarms to agree on shared global states. Practical applicability is demonstrated via deployment on low-cost embedded hardware, achieving fully decentralised coordination through peer-to-peer communication. Finally, the thesis also demonstrates the applications of factor graphs to computer vision tasks, developing an uncertainty-aware rotation estimator that produces temporally consistent camera rotation estimates from RGB images.
This thesis demonstrates that factor-graph-based methods provide a lightweight and generalisable backbone for multi-robot systems, by unifying planning, exploration, formation, and consensus under a distributed, probabilistic approach. We envisage that these methods unlock the potential for large, heterogeneous swarms of the future to perform complex, coordinated behaviours in dynamic and uncertain environments.
Version
Open Access
Date Issued
2025-09-20
Date Awarded
2026-03-01
Copyright Statement
Attribution-Non Commercial-No Derivatives 4.0 International Licence (CC BY-NC-ND)
Advisor
Davison, Andrew J
Sponsor
Dyson Technology Ltd
Engineering and Physical Sciences Research Council
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)