Autonomous robotic systems require sensory information to carry out tasks such as mapping, exploration, and object manipulation. One approach for acquiring additional information about the environment is by employing a mobile manipulator and depth sensors. However, determining the most optimal views that yield the maximum information gain presents a significant challenge.
In this work, a scaled-down implementation of a grid-based next-best view (NBV) planner is presented for 3D object reconstruction. This approach focuses on utilising a Deep Q-network (DQN) with a constrained action space and minimal state information, leading to faster training and less complexity. The reinforcement learning approach can effectively find the NBVs in a 2D grid-based world.