Eyes Up: A Reinforcement Learning-Based, Task-Relevant Sensor Selection Framework

Abstract

To complete the task it was designed for, a robot must have the ability to obtain information from the outside world. This ability of perception usually involves rich capabilities, which give the robot a thorough understanding of its physical environment. In practice, robots with rich sensory capabilities become too sensitive to irrelevant environmental information, like background colors or textures. On the other hand, robots with insufficient sensory capabilities can fail to complete the task. I propose a novel reinforcement learning (RL) framework to automatically tune the richness of a robot’s sensory data. The intuition of my framework is to allow a robot to jointly discover control policies and select sensors; ideally, the robot’s elected sensors provide enough task-relevant information to complete a task. Results on a toy problem indicate that my framework finds a successful navigation policy while dropping out sensors that vanilla RL frameworks do not, but only after extensive training and hyperparameter tuning. The problem statement for another task, robot manipulation, is also presented. After a discussion of privacy and robotics, a topic motivated by my proposed framework, I conclude with proposed directions for future work.