Building Skill Learning Systems for Robotics

摘要

Skill-generating policies have enabled robots to perform a wide range of applications as for example assembly tasks. However, the manual engineering effort for such policies is fairly high and the environment is frequently required to be rather deterministic. For expanding robot deployment to low-volume manufacturing two challenges need to be addressed. First, the robot should acquire the skill-generating policy not from a robot programmer but rather from an expert on the task and second, the robot needs to be able to operate in unstructured environments. In this paper we present a learning approach that combines imitation learning and reinforcement learning to provide a tool for intuitive task teaching followed by self-optimization of the system. The presented approach is applied to a dual-arm assembly task using a real robot and appropriate simulation models. Whereas pure imitation learning does not result in an acceptable success rate for the considered example, after 400 episodes of reinforcement learning the robot can successfully solve the assembly task.