Data Homogenization Method for Heterogeneous Sensors Applied to Reinforcement Learning

Abstract

In autonomous navigation and route planning, the data obtained by the different sensors play a significant role. On the one hand, more data will lead to faster learning of the behavioral policy. On the other hand, agents equipped with different sensors will need more computing power to process the data, thus requiring more robust equipment and increasing the cost of implementation. In addition, the complexity of the algorithms increases as different types of data, i.e., data with different structures, have to be synchronized. Therefore, this paper addresses the problem of homogenization and synchronization of data provided by heterogeneous sensors. Furthermore, it presents a novel method of estimating data in order to provide the agent with a 360-degree view of the environment, similar to that provided by a laser. The method’s performance compares the different behavioral policies obtained by different viewing angles of a camera with the policy obtained by a laser. The data obtained from the different viewing angles of each sensor are used in a path planning algorithm, which was designed to use a single 24-scan laser as an input source. The results show that the proposed method is robust since the behavior policies can be reused regardless of the viewing angle with which the sensor (camera) is provided. Furthermore, the proposed novel algorithm achieves an average efficiency of 68% and 94% using a 90 and 360-degree camera, respectively.

Keywords

• Artificial intelligence
• reinforcement learning
• heterogeneous data
• homogeneous data
• point cloud
• laser scan