ViNG: Learning Open-World Navigation with Visual Goals
Notes
~/org-roam/gdrive/Notability/Rl/LLM and RL/ViNG.pdf
Questions
Why use two different functions for Traversability and Relative Pose predictor ?
Summary
Goal
Given a set of trajectories collected on a target platform in a target environment can a mobile robot navigate to a goal specified by a goal image
Motivation
Robotics Point of view
No localization, SLAM(Mapping), or Calibration is required
RL Point of view
Learning policy where goal is changing
Approach
Collect trajectories.
Build image based topological map.
During deployment localize starting and goal position in the map and generate subgoals.
Navigate to the subgoal and repeat.
Topological Map Generation
~/org-roam/gdrive/Notability/Rl/LLM and RL/ViNG.pdf