Recover, Discover, Plan: Learning Skills and Concepts from Robot Failures

In Submission

Bowen Li^1,2, Mayank Mishra¹, Y. Isabel Liu³, Stone Tao⁴, Nishanth Kumar⁵, Alexander Gray², Ruwan Wickramarachchi⁶, Jonathan Francis^1,6, Sebastian Scherer¹, Tom Silver³

¹Carnegie Mellon University, ²Centaur AI Institute, ³Princeton University, ⁴AI2, ⁵MIT, ⁶Bosch Center for AI

arXiv

ReSYNC can automatically invent neural abstractions (predicates) for generalizable long-horizon planning tasks with continuous and high-dimensional states.