Autonomous Two-Stage Object Retrieval Using Supervised and Reinforcement Learning

Humans have been sending tele-operated robots into hazardous areas in an attempt to preserve life for many years. The task they are presented with is often challenging and requires cognitive abilities, that is, the ability to process information in order to apply it to a different situation. In this work, we proposed an autonomous approach employing both supervised and reinforcement learning for hidden-object retrieval in two stages. Stage 1 used both learning methods to find a hidden object whereas stage 2 only used reinforcement learning to isolate it. The method is targeted towards field robots with reduced computational power hence we explored the viability of tabular reinforcement learning algorithms. We used a convolutional neural network (CNN) to interpret the state of the scene from images and a reinforcement learning agent for each stage of the task. The robot was presented with a workspace containing piles of rubble under one of which an object was buried. The robot must learn to find and isolate it. We compared the performance of four reinforcement learning algorithms over 500 episodes and found that Sarsa (λ) and RMax were most appropriate for stage 1 and 2 respectively. The approach allowed a robot to learn to complete a search and retrieval task by interpreting images. This could lead to the deployment of such robots in disaster areas eliminating the need for tele-operated platforms.