AceAction-5M Dataset for Hierarchical Goal Learning
November 4, 2026
AceAction-5M is a large-scale dataset of human gameplay across multiple 3D worlds designed for vision-language-action (VLA) learning. It utilizes a VLM pipeline to link high-level instructions to temporally aligned subgoals and local skill-based keyboard/mouse actions.
HOW THIS AFFECTS YOU
●
builderYou can use this data to train models on long-horizon task execution in simulated 3D environments.
●
researcherThis provides a structured hierarchical benchmark for training agents to compose complex skills.