Deep Reinforcement Learning in Robotics Logistic Task Coordination

Reinforcement Learning is a conventional ap-proach in the robotics field for solving control problems asan alternative to classic control algorithms. By learning tosolve logistics tasks through simulated experience (explorationand exploitation paradigm) and making use of artificial neu-ral networks, Deep RL algorithms showed the capability ofgeneralizing solutions even for situations in which it hadnever been trained. In this paper, we present our resultsof adopting DRL as a high-level control algorithm in twodifferent logistic tasks proposed by the Logistics League at 2017Latin America Robotics Competition. We also show the roboticand computational architectures built over Festo's RobotinoPlatform to accommodate our algorithm and the implicationsof training and fine-tuning it in a computational environmentbefore deploying it to the real robot.