Developing Robot Reaching Skill via Look-ahead Planning

Reaching is a basic ability of robots and the basis for many complex manipulation tasks. It is an indispensable ability for robots to enter the human family. Previous methods mostly rely on an accurate model such as inverse kinematics model to solve this problem. However, in practice, the model is often inaccurate due to errors caused by measurements and calculations. In this paper, we present a multi-stage approach to develop robot reaching skill based on an internal model and look-ahead planning. We do not assume that the internal model is absolutely accurate. In fact, the inadequacy caused by the inaccuracy of the model can be compensated by the iterative planning in the fine tuning process. In the first stage, the robot collects training data through self-exploration. In the second stage, the robot uses the training data to build an internal model that establishes the relationship between proprioception and spatial position. In the final stage, the robot will reach the goal through coarse tuning and fine tuning. It first gets a rough action (joint movement) through the internal model and then performs the action. Then it will use a planning method to continually fine-tune the action until it finally reaches the goal. The effectiveness and the efficiency of the proposed method have been verified in experiments with the humanoid robot platform Baxter.