Embodied visual tracking asks a robot to do two things at once: pick out a specific target from visual observations and control its own motion to keep following that target through an environment. A new arXiv preprint, VLX-Go, takes on this problem with a system trained in multiple stages and then refined with on-policy learning.
The abstract sets up the task by noting that following an isolated person is comparatively simple. That framing suggests the paper is aimed at more difficult tracking situations, though the abstract itself does not specify what those are.
Because the announcement only includes the abstract's opening lines, the article's specific contributions—such as network architecture, training details, or evaluation results—are not described in the available text. The title and framing indicate a focus on robust embodied tracking through staged training and refinement.