Robot navigation system in stochastic environment based on reinforcement learning on lidar data
In this paper, we present an approach ensuring efficient pathfinding by a robotic platform among static and dynamic objects in a stochastic environment. The approach utilizes data from two-dimensional laser scanner (lidar) that are fed to the neural network for reinforcement learning. The network is trained based on a three-dimensional room model that contains static objects such as walls, floor, stairs, and random dynamic objects, moving in this space along paths. The output data of the trained network represent the next move of the robotic vehicle on its way to the destination point in the room. The presented approach enables the robotic platform to reach the destination point, accounting for special features of the obstacle in the path using data that is taken from the two-dimensional lidar within a certain timeframe.
Lidar data; Path finding; Reinforcement learning; Robot navigation