Asynchronous Learning Boosts Multi-Robot Control with Delays

Operating multi-agent systems (MASs) safely in uncertain environments is crucial for cooperative robotics, where external disturbances and inaccurate models can compromise performance. Gaussian process (GP) regression models are often used for their interpretable performance quantification. While MASs benefit from cooperative learning by exchanging local GP inferences, existing aggregation methods often overlook heterogeneous computational delays and varying query points among agents. To overcome these limitations, this research proposes an asynchronous cooperative learning strategy. This new approach explicitly considers prediction accuracy, variations in query points, and the effects of delays. Additionally, a distributed control law based on an adjoint MAS is developed to ensure desired control performance. Simulations involving unmanned surface vehicles validate the effectiveness of the proposed method. The results demonstrate substantial improvements in both learning and control performance when compared to state-of-the-art approaches, highlighting its potential for more robust and reliable multi-robot operations.