Navigation

A robot deciding, from its own sensor readings, how to move itself to a target location or object.

Navigation means an agent deciding where to go and how to get there, then moving itself to the target. Traditional robot navigation combines mapping and localization (SLAM), path planning, and obstacle-avoidance control. In embodied AI, navigation puts more weight on doing this in environments the robot has never seen, relying only on sensors such as a first-person camera. A 2018 consensus report from Peter Anderson and a dozen other researchers grouped navigation goals into three types: point goals (reach given coordinates), object goals (find an instance of a category, such as a refrigerator), and area goals (reach a type of region, such as a kitchen); it also proposed the SPL metric, which credits both success and how directly the path got there. The goal can also be specified another way — in natural language, as in vision-language navigation (VLN), or with a picture, as in image-goal navigation. Combining navigation with manipulation gives mobile manipulation.

ExampleA robot given the goal “find the refrigerator” walks through an apartment it has never visited, looking as it goes, and counts as successful once it stops within a distance threshold of the fridge — the report suggests twice the robot's body width.

Also called

Embodied Navigation, Visual Navigation

Related

Vision-and-Language Navigation · Object-Goal Navigation · Point-Goal Navigation · Success weighted by Path Length · Simultaneous Localization and Mapping · Mobile Manipulation

Sources

On Evaluation of Embodied Navigation Agents (Anderson et al., arXiv:1807.06757)