A DRL Approach for Object Transportation in Complex Environments

Robots capable of transporting objects are suitable for many applications with societal and economic impact, such as waste retrieval, disposal, and object manipulation in space or the deep sea. However, formulating a coherent action plan is not trivial due to the size of the search space and the object's physical properties. With the recent advances in Deep Reinforcement Learning (DRL), in this work, we propose, implement, and deploy value-based Deep Reinforcement Methods to tackle the determination of high-level actions that form robust strategies combined with a Probabilistic Roadmap (PRM) method for object transportation through complex environments. The solution was evaluated in a simulation environment and deployed into a real robot. Our results show that DRL can learn strategies effectively, and the robot was able to accomplish its task.