Muthukumar Pandaram

This doctoral thesis aims to explore a key question: how can embodied AI agents, like robots, learn low-dimensional abstract representations or structures from high-dimensional perceptual data for perceptual decision making and action? One can call these abstract representations or structures as inductive biases, priors or constraints. Humans are known to learn task relevant goal-directed representations through different neural mechanisms. This research wants to draw inspiration from how the human brain creates sensorimotor representations and uses these abstract representations for learning to make decisions under uncertainty. I believe such representation learning can help in better generalisation of robot learning to out-of-distribution domains.