Reinforcement learning for robotic assisted tasks

The increment of dependant people for the realization of Activities of Daily Living is a fact that comes implicit with the increase of population and their lifespan given by the modern technological progress. For this reason, Assistive Robotics is a field that is currently researched and gradually implemented. In this project, two different algorithms to train robots to perform those tasks are being studied and compared, Reinforcement Learning and Imitation Learning. The main objective is to learn the benefits, drawbacks and suitability of both for a given task. To do the experimentation, the Assistive Gym software will be used, to train the nets with Reinforcement Learning and to perform the movements with Dynamic Movement Primitives learned from a demonstration. The environment will consist in a Sawyer robot cleaning the right arm of a human lying down on a bed. The potential of each methodology will be tested in different scenarios, having the user’s arm straight or partially flexed. Furthermore, small implementations to improve the execution of the task will be implemented for the Dynamic Movement Primitives to test the performance and the similarity to the demonstration. Both methodologies are capable to perform the desired task successfully, the model trained with Reinforcement Learning is more suitable for an individual user that will use the robot in a long term, due to the training time. The Imitation Learning algorithm is faster to teach and its trajectory and velocities are easily adaptable to the preferences of the user, it is more suitable in places where there is a high rotation of users that require urgent and short attention, such as clinics or hospitals