News

Abstract: AI in the digital domain (i.e, text and images) is progressing at a dizzying pace, in large part due to an abundance of data and an increasing reliance on a larger compute budget. However, AI in the physical domain significantly lags behind. In fields such as engineering, healthcare, robotics, or economics data tends to be scarce, difficult to label, or challenging to aggregate due to privacy concerns. Moreover, while digital AI systems can make mistakes without real-world consequences, an error made by a physical AI system can cause harm, death, or financial loss. In this talk, I discuss strategies for improving the performance and robustness of physical AI systems. I present techniques for solving difficult physical perceptual tasks, ideas for addressing the small data problem, and methods for constructing agents that can perform reliably when the cost of an error is high. Additionally, I discuss successful deployments of these systems across a variety of fields, and present ideas to make physical AI systems more robust in the future.