Robust and High-Precision End-to-End Control Policy for Multi-stage Manipulation Task with Behavioral Cloning

Abstract

In this paper, we propose a multi-stage task learning method that trains an end-to-end policy to control a 6-DoF robot arm to accomplish pick-and-place operation with high precision. The policy is mainly composed of CNNs and LSTMs, directly mapping raw images and joint angles to velocities command. In order to acquire a robust and high-precision policy, several techniques are introduced to boost performance. Augmentation trajectories are designed to alleviate compounding error problem, and dataset resampling is used to solve imbalanced data issue. Moreover, Huber loss for auxiliary outputs is illustrated to be very effective in multi-objective optimization problems, especially in robot learning field where sample complexity needs to be reduced desirably. To verify the effectiveness of our method, experiments are carried out in Gazebo simulator with UR5 arm and Kinect v1 camera. Our visuomotor policy can achieve a success rate of $$87\%$$ on the pick-and-place task. The results of our experiments demonstrate that, with the skills we mention, behavioral cloning can effectively help us to learn good visuomotor policies for long-horizon tasks.