Announcement
RAX aims to turn robot exploration into reusable skills
A post describing RAX reports a 78.3% RoboDojo result, versus 24.9% for the prior state of the art.
TLDR
In a series of posts, the author says RAX uses depth and force observations to explore robot tasks, then saves successful perception, planning and control routines as programs it can reuse across sessions. Unfamiliar situations return control to the agent for further exploration. The author reports 78.3% on RoboDojo, versus the prior best of 24.9%, and 46.5% on BEHAVIOR-1K, versus 8.2%.
Combined views
1.2K
6 Sources, first seen ago
