Distributed Deep Reinforcement Learning for Dynamic Task Scheduling in Multi-Robot Systems

With the increasing task scale and dynamic complexity of multi-robot systems in automated production lines, we design a dynamic distributed task scheduling framework to accelerate the convergence of the combinatorial optimization model, which is suitable for time-varying multi-task and multi-robot scenarios. By integrating an empirical learning model and a teammate collaboration model, a distributed deep reinforcement learning algorithm is formulated with a limited number of workstations. Each workstation is designed as an independent agent, interacting with the environment to learn the current allocation state of robots. Then, a Q-learning network is trained to extract high-dimensional features from the state and optimize the task scheduling policy. Besides, a greedy strategy is incorporated with the Q-learning network to favor actions that show an increasing trend in Q-values, which enables the algorithm to prioritize higher-priority tasks when facing resource limitations. Simulations with three different workload intensities demonstrate that our algorithm achieves a respective enhancement in overall performance of 3.50%, 8.16%, and 3.86% compared with the foundational deep reinforcement learning models.