Real-time Obstacle Avoidance in Robotic Manipulation Using Imitation Learning

We propose a novel trajectory planning algorithm to avoid obstacles in robotic manipulation by using imitation learning method. It focuses on how to plan feasible trajectories in the manipulation environment by imitating human experience. The main components of our algorithm include a path point prediction network and a trajectory generation strategy. The network is primarily composed of several Long Short-Term Memory (LSTM) layers and a Mixture Density Network (MDN) layer with Gaussian functions, thus it can cope with sequential information and fit the multimodal dataset well. To improve the smoothness of the trajectories generated by the networks, trajectory points are sampled from the Gaussian function which has the minimal change in the configuration space. Besides, multiple trajectories are generated for a given input and the best one can be selected to accomplish the task by improving the precision of the planning algorithm. Finally, simulation experiments conducted in Gazebo simulator verify that our planning algorithm has good performance in robotic manipulation with obstacle avoidance.