Zero-shot coordination

Aim

To enhance the coordination capabilities of autonomous agents during the deployment, either with other autonomous platforms or with human operators.
 
Objectives

1.    Develop novel algorithms for zero-shot coordination
2.    Demonstrating the validity of these methods in real-robotic platforms
 
 Description

There is a growing interest in developing teams of robots capable of coordinating seamlessly to accomplish complex tasks. This coordination, though challenging, introduces new possibilities for collaborative missions. Existing research in multi-agent systems, particularly in multi-agent reinforcement learning (MARL) [1], has predominantly focused on scenarios where cooperation is achieved via co-training with a set of teammates. However, co-training has several disadvantages. The policies developed within the team, are typically not transferable outside of the team, as they depend on conventions that are spontaneously generated during training. As such, these policies are not good at generalizing to agents outside the team, and performance degrades greatly. This particularly affects humans, as the dependency on conventions makes the behaviours of these agents hard to understand, which limits the deployment in real-world scenarios. On the other hand, zero-shot methods seek agents that are capable of coordinating their actions with teammates outside of their training team and which exhibit different or unexpected behaviours. In this way, the resulting agents are able to cooperate more generally, which allows better coordination with other agents, including humans [2].    
   
This proposal seeks the advancement of autonomous agents capable of zero-shot coordination. This research holds significant potential for enhancing collaboration in real-world robotics. Our emphasis is on leveraging reinforcement learning techniques [3] to achieve autonomous agents capable of zero-shot coordination on realistic robotic platforms. By tackling the inherent challenges associated with zero-shot coordination, this research aims to provide innovative insights and practical solutions applicable to autonomous robotics across industries. The overarching goal is to elevate the capabilities of robotic teams, enabling them to collaborate efficiently and adaptively in dynamic real-world environments.
 
 References

[1] Albrecht, S. V., Christianos, F., & Schäfer, L. (2024). Multi-agent reinforcement learning: Foundations and modern approaches. MIT Press.
[2] Mirsky, R., Carlucho, I., Rahman, A., Fosong, E., Macke, W., Sridharan, M., P. Stone, & Albrecht, S. V. (2022, September). A survey of ad hoc teamwork research. In European conference on multi-agent systems (pp. 275-293). Cham: Springer International Publishing.
[3] Sutton, R. S., Barto, A. G. (2018). Reinforcement Learning: An Introduction. The MIT Press.  
 

Research theme: 

Autonomous sensing platforms

Principal supervisor: 

Dr Ignacio Carlucho
Heriot-Watt University, School of Engineering & Physical Sciences
ignacio.carlucho@hw.ac.uk

Dr Stefano Albrecht
University of Edinburgh, School of Informatics
s.albrecht@ed.ac.uk