Learning Manipulation Trajectories Using Recurrent Neural Networks

Robots assisting disabled or elderly people in the performance of activities of daily living need to perform complex manipulation tasks which are highly dependent on the environment and preferences of the user. In addition, these environments and users are not suitable for the collection of massive amounts of training data, as the manipulated objects can be fragile, and the wheelchair-bound users might have difficulty recovering from a failed manipulation task. In this paper, we propose an end-to-end learning mechanism for the type of complex robot arm trajectories used in manipulation tasks for assistive robots. The trajectory is learned using a recurrent neural network that can generate the trajectory in real-time based on the current situation of the end-effector, the objects in the environment and the preferences of the user. The learning data is acquired from a simulation environment where the human can demonstrate the task in a simulation closely modeling his or her own environment. Experiments using two different manipulation tasks show that the robot can learn the manipulation planning as well the ability to recover from failure.