Description

Traditionally, robots and autonomous vehicles are programmed to follow rules to achieve their goals. This is costly because each rule has to be written by an engineer. It’s also inflexible because small changes in the task requires that a new set of rules be written. Such classical artificial intelligence is increasingly being complemented with solutions from machine learning (ML) – the program actually learns from experience. Machine learning has continued to improve in performance benchmarks across all domains, yet it has not been fully exploited. For this reason we believe that achieving effective solutions requires cross-disciplinary efforts involving traditional engineering, and a fast development cycle that incrementally incorporates state of the art machine learning methods as they emerge. Since machine learning requires a lot of data, we rely heavily on simulation to generate some of that data. We utilize components and inspiration from stochastic control, probabilistic models, deep reinforcement learning, neuroscience and physics.