Computer Vision News - June 2022

11 Is Mapping Necessary for Realistic... The problem of goal-navigation was the subject of a previous article in Computer Vision News of December . We are now again looking into this from a different perspective: How is Realistic PointGoal Navigation different from idealized one, and howmuch does it rely on Mapping? PointNav is a navigation task where an Agent is initialized in a previously unseen environment and is tasked to reach the goal specified relative to its starting location. The action space is discrete and consists of four types of actions: stop (to end the episode), move forward by 0.25m, turn left and turn right by a specified angle. In this work, the agent was equipped with an RGB-D camera mounted at a height of 0.88m and tilted -20°. Camera’s resolution was 360x640 pixels with 70° horizontal field of view and base radius of 0.18m. PointNav comes under two versions: v1) idealized setting: the agent is equipped with noise-free camera and access to ground-truth localization and movement is deterministic. In idealized setting, with no noise, map-less navigation models trained with large-scale reinforcement learning achieve 100% success. State-of-the-art approaches seem to have solved this problem; v2) realistic setting, where the agent must deal with actuation and sensing noise, and lack of high-precision localization in indoor environments. This is