This project focuses on distributed cooperation and decision-making in multi-agent robotic systems, with a particular emphasis on Multi-Agent Reinforcement Learning (MARL) in partially observable, spatio-temporal environments. The research investigates how teams of autonomous agents can learn coordinated sensing and control policies that enable efficient exploration and long-term autonomy under uncertainty.

A central component of the work is spatio-temporal environmental modelling using Gaussian Process regression, which provides probabilistic representations of unknown and dynamic environmental processes. These models are updated online from sparse, local observations and are explicitly integrated into the agents’ decision-making pipelines. The problem is formulated under partial observability, where agents must act based on incomplete and decentralized information.

The research develops MARL-based coordination strategies that balance exploration and exploitation, learn cooperative behaviours, and scale to increasing team sizes while remaining communication-efficient. Learning is performed in a distributed manner, enabling agents to adapt their policies as the shared environmental model evolves.

The proposed methods are validated on heterogeneous robotic teams, including unmanned aerial vehicles (UAVs) and ground robots, demonstrating the advantages of learned cooperative policies over single-agent and heuristic approaches. Potential application domains include environmental monitoring, precision agriculture, infrastructure inspection, and disaster response.

This project focuses on distributed cooperation and decision-making in multi-agent robotic systems, with a particular emphasis on Multi-Agent Reinforcement Learning (MARL) in partially observable, spatio-temporal environments. The research investigates how teams of autonomous agents can learn coordinated sensing and control policies that enable efficient exploration and long-term autonomy under uncertainty.

A central component of the work is spatio-temporal environmental modelling using Gaussian Process regression, which provides probabilistic representations of unknown and dynamic environmental processes. These models are updated online from sparse, local observations and are explicitly integrated into the agents’ decision-making pipelines. The problem is formulated under partial observability, where agents must act based on incomplete and decentralized information.

The research develops MARL-based coordination strategies that balance exploration and exploitation, learn cooperative behaviours, and scale to increasing team sizes while remaining communication-efficient. Learning is performed in a distributed manner, enabling agents to adapt their policies as the shared environmental model evolves.

The proposed methods are validated on heterogeneous robotic teams, including unmanned aerial vehicles (UAVs) and ground robots, demonstrating the advantages of learned cooperative policies over single-agent and heuristic approaches. Potential application domains include environmental monitoring, precision agriculture, infrastructure inspection, and disaster response.

Related Publications