Learning of Robotic Throwing at a Target using a Qualitative Learning Reward

Autonomous learning and adaptation of actions is critical for robots to operate in unstructured, everyday environments. Reinforcement learning (RL) methods are often applied for this. However, efficient RL requires the determination of an appropriate reward function, which is a complex problem even for domain experts. In this paper we investigate if a standard robotics reinforcement learning method called PoWER can be effectively utilized with a simple, qualitatively determined reward, instead of with a complex reward function. Our use-case example is robotic throwing at a target. However, for increased complexity, we perform throwing with a 7 degree of freedom arm of a humanoid robot, and a two-dimensional target space, i.e., the target is placed arbitrarily on the plain in front of the robot, which needs to learn the direction and the distance. Results show that learning with a simplified reward function that practically assigns a qualitative reward, just as a person would, can still be effectively used for RL using PoWER.