Abstract In multi-robot collision avoidance navigation problems, it is common to encounter situations where robots get stuck and cannot avoid collisions or complete navigation tasks, known as deadlock problems. There are various reasons for the occurrence of deadlock problems, such as issues with the navigation method itself. Optimal reciprocal collision avoidance (ORCA) is widely used in multi-robot navigation. However, in some scenes, such as where robots distribute symmetrically, ORCA usually fails to plan reasonable collision avoidance velocities, resulting in the robot getting stuck. In this paper, we aim to address the deadlock problem caused by ORCA. We first analyze and simplify the planning process of ORCA from the principal horizon to break the deadlock situation. Then, deep reinforcement learning is used to optimize the escape velocity and collision avoidance responsibility of ORCA to guarantee collision-free navigation. We compare our method with the original ORCA and other learning-based methods. The results show that the proposed method is beneficial in reducing the deadlock probability and the collision probability in multi-robot navigation.
Liu et al. (Thu,) studied this question.