This demonstration shows an embodied agent built with active inference principles performing complex household tasks in the Habitat-Sim environment. The agent locates hidden objects, reasons about their states, tracks beliefs, and executes an 18-step plan to make and serve tea. It jointly reasons about multiple objects and uses semantic priors and negative evidence. The project includes interactive visualizations of the agent's beliefs, policies, and hidden states.
In the AI & ML space, Active Inference World Models takes a focused approach. It focuses on demonstrating how active inference can be used to build practical world models for embodied agents that reason under uncertainty. It is built as an open-source project for AI researchers. Active Inference World Models costs nothing to use. Active Inference World Models is available on the web.
CPNS Lab builds and maintains Active Inference World Models, and it first shipped in 2026. Among its 5 catalogued features are Belief Tracking, Multi-Object Search, and Hidden State Inference.
Summary written by a language model from the project’s public pages.
What PulseGate has recorded for this listing
Same category — not a similarity match