A STUDY OF REWARD FUNCTIONS IN REINFORCEMENT LEARNING ON A DYNAMIC MODEL OF A TWO-LINK PLANAR ROBOT

摘要

Reinforcement Learning is one approach for executing a point-to-point movement using complex multi-link robots in an environment. The work presented in this paper, we: (1) use torque commands as input for a two-link manipulator; position commands do not take the dynamics of the system into consideration, (2) identify the optimal composition of a set of reward functions in fulfilling certain requirements. These requirements include: a smooth velocity profile while reaching the target, no overshooting and a minimum number of steps. We experimentally validate our approach on a two-link planar robot for reaching different target positions.