Image-Goal Navigation

Giving a robot a photo of a destination and letting it find its own way there in an unfamiliar space.

Image-goal navigation is a benchmark task in embodied navigation: an agent is placed in a previously unseen indoor environment and given only a photo taken at the goal location, then must reach that spot using nothing but its own camera. Early work includes Zhu and colleagues' 2016 target-driven visual navigation, which also introduced the AI2-THOR simulator; platforms such as Habitat later turned the task into a standard benchmark. The original setup has two weaknesses: a randomly captured goal image can be as uninformative as a blank wall, and the goal photo must be taken with camera settings matching the robot's own camera. To address this, Krantz and colleagues proposed Instance-Image Navigation (InstanceImageNav) in 2022, where the goal image targets a specific object and can be taken with any camera, making it closer to real-world use.

ExampleA user photographs a chair in their study on their phone and sends it to a home robot, which explores the house and stops next to that chair.

Also called

ImageNav, Image Navigation, ImageGoal Navigation

Related

Navigation · Object-Goal Navigation · Point-Goal Navigation · Vision-and-Language Navigation · Habitat · AI2-THOR

Sources

Target-driven Visual Navigation in Indoor Scenes using Deep Reinforcement Learning (Zhu et al.)
Instance-Specific Image Goal Navigation: Training Embodied Agents to Find Object Instances