Object-Goal Navigation

Telling a robot only an object category, such as 'chair,' and having it find one on its own in an unfamiliar house.

Object-goal navigation is one of the standard tasks in embodied navigation. The 2018 navigation-evaluation working-group paper split navigation goals into point goals, object goals, and area goals, and the 2020 paper “ObjectNav Revisited” unified the detailed evaluation protocol. In this setting, the agent is placed at a random position in a previously unseen indoor environment and given only an object category name, then has to find an instance of it using a first-person RGB-D camera as it moves. The Habitat 2020 Challenge specifies that the episode counts as successful if the agent actively issues a “stop” action within 1 meter of any instance of that category and the object is visible from the stopping point; common metrics are success rate and success weighted by path length (SPL). It is harder than point-goal navigation because the agent must recognize the object, infer where it is typically located, and explore efficiently.

ExampleIn the Habitat simulator, an agent given the instruction “find the toilet” starts in the living room, infers the toilet is most likely in the bathroom, explores toward it, and stops within 1 meter once it finds one.

Also called

ObjectNav

Related

Navigation · Point-Goal Navigation · Image-Goal Navigation · Vision-and-Language Navigation · Success weighted by Path Length · Habitat

Sources

ObjectNav Revisited: On Evaluation of Embodied Agents Navigating to Objects
Habitat Challenge 2020
On Evaluation of Embodied Navigation Agents

As of

2020-06