Actor-critic based on attention model for multi-robot collaborative backend optimization

Abstract Backend optimization is an essential component of simultaneous localization and mapping (SLAM).Collaborative backend optimization in multi-robot systems refers to the process of extending single-robot collaboration optimization to coordinate and optimize the backend processes of multiple robots working together, enhancing overall system performance and efficiency.In this paper, a deep reinforcement learning model based on attention mechanisms called MAS-AA specifically tailored for collaborative backend optimization in heterogeneous multi-robot systems is proposed, solved the problem of heterogeneous multi robot system collaborative backend optimization not being able to optimize the selection of map points and pose nodes based on constraints between map points and pose nodes considering robot states and attributes. Firstly, we introduce a collaborative attention neural network designed for multi-robot back-end optimization, along with a collaborative decision-making neural network based on deep reinforcement learning. Secondly, we delve into an optimization mechanism based on the optimal collaborative chain, as well as a multi-robot bundle adjustment algorithm derived from this mechanism. Lastly, we design and implement a cost function for the decision-making model based on collaborative attention, as well as a reward function for the collaborative model. We further present a learning methodology that combines the weight update processes of both neural networks.Simulation experiments validate the significant enhancements achieved by our algorithm in terms of localization accuracy and mapping performance in multi-robot collaborative backend optimization. Effectively addressing the limitation of improvement in collaboration performance caused by the inability to perceive subsequent collaboration states in the application of attention models in multi-robot collaborative backend optimization.