Learning to Act in Interactive Worlds.
AceAction is a research initiative exploring how intelligent agents understand goals, perceive dynamic environments, and translate intentions into actions. Our interests span vision-language-action (VLA) learning, action-conditioned world modeling, and long-horizon decision-making in interactive 3D worlds.
Hierarchical Goals for Generalist Action Learning Across 3D Game Worlds
A multi-game action-learning dataset with hierarchically structured goals, subgoals, and skills across 42 games and over 5 million annotated episodes.
Abstraction-Aligned Action Learning for Vision-Language Agents in 3D Game Worlds