Deep Learning-Based Fast Inference of Team Performance in Heterogeneous Multi-Robot Systems

Abstract The effectiveness of multi-robot teams depends on various factors, including the size of the team, the capabilities of each individual robot, the synergies between them, and the coordination strategies employed. However, current studies on multi-robot task allocation often assume pre-established teams, which results in a focus on optimizing resources rather than designing teams from a diverse range of robots or selecting the most suitable team from a set of potential configurations. Our research focuses on predicting a team’s performance in a given mission based on its composition. This task has significant implications for selecting or designing multi-robot teams, as it helps to optimize performance while minimizing costs. To this end, we propose a novel fast inference method for estimating team performance using supervised learning in multi-robot task allocation situations. We considered a general multi-robot task allocation issue for spatially distributed missions. For evaluation purposes, we used a large dataset of varied problem instances, two different solvers, and five different team performance metrics derived from the solutions computed by these solvers. The results demonstrate that long short-term memory recurrent neural networks, aided by convolutional layers, can accurately predict team performance metrics such as makespans and intra-team communication metrics derived from the solutions obtained by the two solvers. We also highlight the efficacy of transfer learning in fine-tuning models from a simpler sub-optimal solver to a more complex, optimal solver, demonstrating improved accuracy with reduced training set size and computational burden.