This talk addresses the development of robotic systems that can collect data in the natural world, either on land or underwater. While the last decade has seen immense progress in the development and deployment of self-driving vehicles, moving around in the natural world without a map and outside of cities remains much more challenging. Most of the world remains undeveloped, and understanding and describing the environment is a huge challenge which is critical for many reasons. Navigation in dynamic unstructured environments is still an area where there is limited consensus over the correct kinds of approach, partly due to the diversity of environments, as well as need to learn`on the fly’. In this talk ,Greg will discuss the key challenges, and present solutions that allow autonomous vehicles to drive in the forest and underwater using vision-guided machine learning. In this body of work, a key theme is the synthesis of model-based methods that can exploit powerful prior knowledge, and model-free methods that can adapt to the variations in an unconstrained world.