Announcement
SOLE-R1 accepted to NeurIPS 2026
A SOLE-R1 researcher says its zero-shot reward predictions served as the sole signal for online reinforcement learning on more than 20 unseen robot-manipulation tasks.
TLDR
A SOLE-R1 researcher announced the project's acceptance to NeurIPS 2026. The researcher says robots learned more than 20 unseen manipulation tasks using zero-shot reward predictions as their sole learning signal. They started with randomly initialized policies and a 0% success rate, without demonstrations or ground-truth rewards.
Combined views
3.2K
2 Sources, first seen 8h ago
