Reinforcement learning for multi-agent systems under semantic and perceptual uncertainties

Multi-agent collaborative tasks for autonomous agents such as search and rescue tasks often take place in environments where properties and dynamics of the environment are unknown or uncertain. Often, such collaborative tasks are temporal high-level tasks, possibly with explicit time constraints, and the agents’ sensors used to perceive the properties of interest suffer from uncertainties. Hence, challenges arise when attempting to perform the specified task and simultaneously collaboratively learn the environment’s and targets’ uncertainties under potentially unreliable communication channels.

In this project, these challenges are addressed by developing robust algorithms that inherently take these limitations into account. This is achieved by leveraging multi-agent, uncertainty-aware collaborative reinforcement learning together with decentralized sensor fusion and goal-oriented communication, enabling successful task execution of heterogeneous collaborating autonmous agents such as unmanned vehicles.