Meta released InteractiveWorld, a simulator for wearable AI assistants. It lets a model see through a person’s eyes and give commands. A human model then acts in a 3D world and the goals are checked against world state. The repo includes 44 scenes and 120 frozen tasks. It supports 36 action families for movement, manipulation and tool use. A Python API and command-line tools are provided. The benchmark evaluates assistants by pairing them with a fixed human model. Tasks span six categories including guidance, planning and accessibility. Setup requires Isaac Sim, OmniGibson and BEHAVIOR-1K assets. Human model files must be downloaded separately under non-commercial licenses.
What it unlocks
Building and testing interactive AI assistants in a verifiable 3D human-centric simulation.
- 120 multi-turn tasks
- 44 scenes
- 36 action families
- 6,831 BEHAVIOR objects
What you need to act on it
- Linux with an NVIDIA GPU
- 100 GB disk space
- BEHAVIOR-1K dataset
- MANO and AMASS model files
- non-commercial research license
Sources