Offline Imitation Learning for Robotic Control using Contrastive methods

Autonomous robotic applications have a huge scope of application in a variety of sectors including manufacturing, warehousing and personal robotic systems. Training a robotic system to autonomously perform a specified set of tasks is a key element in this. Deep learning has proved to be the most efficient tool to train a robotic system. Current methods for using deep learning for robotics involve, Generative Adversarial Networks that employ two networks, the generator and the discriminator. All of the current algorithms require access to a good simulated environment to train. This is a severe limitation as many tasks do not have access to a good simulator. We propose an offline imitation learning method for autonomous robotic tasks, that involves logging data from the environment in the training phase and evaluating it for performance in the environment. In our proposed method, we additionally use Mean Squared error between generated action and logged action for training the generator network which provides an additional signal for the network to learn. With current trends of increasingly large amounts of data collection, our data-driven approach holds enormous potential to outperform the current methods in terms of efficiency.