Generally capable agents emerge from open-ended play

Abstract

In new work, algorithmic advances and new training environments lead to agents which exhibit general heuristic behaviours.