Robotic vision is embodied, constrained and discriminative. Embodied: cameras operate on a physical agent that interacts with its environment. Constrained: robots have finite resources including energy and time. The environment and application present additional challenges. Discriminative: vision drives decisions: "Should I go left or right?" "Is it okay to move that chair?" "Is that a real apple?"

Nature has not settled on a single eye design: different evolutionary niches benefit from different combinations of optics and computation. Robots similarly benefit by matching applications with appropriate embodiments, sensing and compute.

Learning to See with New Eyes

Machine learning has taken massive strides with conventional cameras, but important opportunities arise in learning new ways of seeing. How can machine learning automate the design, manufacture, and interpretation of new kinds of camera?

This emerging field jointly designs optics and algorithms to deliver better images. Recent years have seen a tantalizing profusion of new imaging technologies.

From Photons to Actions

We are exploring the union of computational imaging and robotic vision. This work will establish the key principles by which cameras and algorithms are designed to enable the next generation of robotic autonomy.