Implementation of Learning to Enhance Cost Reduction in Object Search

Abstract

Many object search algorithms in unknown household environments search greedily, since many algorithms, unaware of where task-critical objects may be located, naively search nearby containers until the target object(s) can be found, thus making them problematically inefficient. Efficient object search in unknown environments requires predicting what lies beyond the known space. In our work, we demonstrate that machine learning is an effective tool for making such predictions about where unseen objects are likely to be found in household environments. In our research, we trained and tested neural networks that we designed to complete complex tasks in household environments while minimizing the overall cost of running a certain algorithm. We train a graph neural network in similar household environments and then use the resulting model to predict the probability of the object’s existence at unseen locations. By using our learned models, the algorithm can anticipate where target object(s) are likely to be found and allow the robot to make more informed decisions, even in unfamiliar situations, enabling the system to adapt its decisions based on training done in similar environments. We finally report the learned model that results in the best object-search performance.