The invention discloses a
mobile robot autonomous
obstacle avoidance method based on multi-thread asynchronous deep
reinforcement learning, and the method specifically comprises the steps: adding a multi-task asynchronous parallel mechanism on the basis of a PPO
algorithm, constructing MAPPO, separating different
obstacle avoidance task scenes, and training the scenes at the same time; the method comprises the following steps: on the basis of nokov-
lidar multi-sensor sensing information, constructing a
robot environment state observation space; designing a discretized action space based on the
global grid world navigation map, and setting
kinematics constraints for state updating; designing a navigation reward function, and guiding the
mobile robot to make an optimal
obstacle avoidance decision in a complex environment; an early collision prediction module is established based on a multi-layer
perception mechanism, collision information from a perceptible environment is deduced, and an optimal obstacle avoidance strategy is trained in combination with MAPPO learning. According to the invention, sufficient
mobile robot-environment interaction can be realized, the exploration capability of the
robot action
decision model is improved, and real-time obstacle avoidance of the
robot in the process of moving to the target is ensured.