Point-Goal Navigation

Giving a robot a goal coordinate relative to its start point and having it navigate there on its own in an unfamiliar environment.

Point-goal navigation is the most basic embodied navigation task, defined in the 2018 navigation-evaluation working-group paper: the agent starts in a previously unseen environment, and the goal is given as a coordinate relative to the starting point, such as “5 meters north, 3 meters west” — no object recognition is needed, only obstacle avoidance, planning, and reaching the point. The Habitat 2020 Challenge specifies that issuing the stop action within 0.36 meters of the goal, twice the agent's body radius, counts as success, measured by success weighted by path length (SPL) to capture how efficiently the agent got there. Erik Wijmans and colleagues' DD-PPO (2019), trained on 2.5 billion steps of simulated experience, essentially “solved” this task when given an RGB-D camera plus GPS and compass; the 2020 challenge then removed GPS and compass and added sensor noise to bring the task closer to a real robot.

ExampleGiven the goal “5 meters north, 3 meters west” in Habitat, the agent, relying only on a first-person RGB-D camera, navigates around a sofa and a wall and stops within 0.36 meters of the goal to succeed.

Also called

PointNav

Related

Navigation · Object-Goal Navigation · Success weighted by Path Length · Habitat · Image-Goal Navigation · Visual Odometry

Sources

On Evaluation of Embodied Navigation Agents
Habitat Challenge 2020
DD-PPO: Learning Near-Perfect PointGoal Navigators from 2.5 Billion Frames

As of

2020-06