Configuration-Aware Robotic Trajectory Planning Based on Deep Deterministic Policy Gradients for Active Object Tracking

Active object tracking provides a solution simultaneously addressing tracking and camera control, which can be usually implemented by robotic manipulators amounted with the camera at the end effector (e.g., move left, move forward, etc.). Trajectory planning is particularly the key technique for the control of robots during the tracking process. However, traditional trajectory planning methods are unable to achieve active tracking using robotic manipulators with different configurations, leading to poor generality and long training times. To address these problems, this paper proposes an improved trajectory planning method based on deep deterministic policy gradients (DDPG), to implement active object tracking tasks. A plug-in is designed to comply with the configuration of robotic manipulator. To improve generality and training efficiency, a novel reward function is proposed. Several experiments are conducted using a two-DoF robotic manipulator and a UR5 robot in the simulation and the physical two-DoF pantilt-zoom (PTZ). The proposed method can be applied to robots with different configurations. Experimental results demonstrate that the proposed method is feasible and has faster training speed and better active tracking performance compared with PPO, TD3, and SAC.