Ananth Rachakonda

Robots learning intelligent behaviors through Reinforcement Learning are notably sample-inefficient, requiring extensive interaction with the environment while often failing to generalize in new ones. These shortcomings, though insignificant in simulation, make their real-world application impractical and unreliable. The governing physical laws of the robots and their environments, though many dimensional and task dependent in theory, could be resolved into simpler abstractions of field functions, state representations and motion priors which would form the intuitive physics model of the robot. We posit that endowing robots with such intuitive physics could significantly enhance their sample efficiency while aiding safe and meaningful exploration of their environments. This will be the driving motivation of the doctoral dissertation.