Research

Robots have become an integral part of today’s community as they are ubiquitous, performing numerous tasks in countless real-world scenarios. Along with real-world robots, there has been a tremendous increase in the development of photo-realistic simulators that could be used to foster training and learning of AI agents.

Traditionally, agents are trained by interacting with these simulators or real-world, which typically takes a long duration of timesteps. Especially in real-world scinarios, its desirable to have an agent quickly learn a task and adapt to ever changing world.

By using all the data obtained from the interactions of the agents with the real-world and simulators, we can construct algorithms to extract primitive skills or behaviours from numerous robotic tasks. With the recent advances in distributed computing and parallelization, this process could be made even more efficient.

Methods which use these primitives tremendously speed up agent training. They could be used to quickly learn and adapt to different novel downstream tasks using prior experiences.