Active Inference World Models | CPNS Lab

Skip to content

Embodied AI · Active inference · World models

Find the mug.<br>Make the tea.

The first demonstration showed an agent using semantic priors and negative evidence<br>to locate a hidden mug. The new Habitat-Sim task carries the same reasoning into<br>multi-object search, hidden-state inference and an 18-action plan that ends with<br>tea served at the dining table.

Watch the new demo<br>See the original search demo<br>Open the preprint

objects reasoned about jointly

18<br>dependent task actions

Visible<br>beliefs, policies and hidden states

New embodied demonstration<br>From search to structured action

The agent finds the mug, teabag and kettle, checks whether the kettle contains<br>water, fills and boils it, pours, steeps and serves.

Habitat-Sim · ReplicaCAD · active inference

Watch the full reasoning trace →

New demonstration · July 2026

From uncertain search to making a cup of tea

The agent must locate three objects, reason about what is currently known,<br>satisfy action dependencies and track state changes from an empty kettle to a<br>served cup of tea.

Multi-object task · full decision trace

Your browser does not support embedded video.<br>Open the MP4 instead.

The main view shows the agent in the task environment; the inset shows its<br>first-person observation. The right-hand panels expose location beliefs,<br>hidden-state confidence and the task graph while the lower panel shows the<br>selected action and progress.<br>Read the Technical Note.

Decision trace JSON<br>Download video

00:00<br>Evaluate the search space<br>Candidate object–location pairs are ranked by probability, information gain, distance and task relevance.

00:12<br>Opportunistic discovery<br>The kettle is found while the current task frontier is still to locate the mug.

00:28<br>Negative evidence<br>Empty locations reduce their posterior probability and redirect the next inspection.

00:56<br>Objects located<br>The search phase resolves the mug, teabag and kettle before task execution proceeds.

01:20<br>State-dependent actions<br>The agent checks, fills and boils the kettle only when dependencies and belief preconditions are met.

01:39<br>Pour, steep and serve<br>Action effects update the hidden-state beliefs until the task goal is satisfied.

01<br>Multi-object inference

Beliefs are maintained over the locations of the mug, teabag and kettle rather than a single target.

02<br>Task relevance

The search policy can value an object that is useful downstream, even when another search node is currently active.

03<br>Hidden object states

The model tracks whether water is present and hot, whether tea is brewed and whether it has been served.

04<br>Dependent action sequence

Eighteen graph-linked actions connect locating and moving objects to filling, boiling, pouring, steeping and serving.

Decision-making logic

One model decides what to inspect and what to do next

Search and task execution are coupled. Evidence changes beliefs; beliefs change<br>which actions are available and valuable; successful actions change the world<br>state that the next decision is conditioned on.

Rank object–site hypotheses

The policy compares candidate searches across all unresolved task-relevant<br>objects, balancing current probability, expected information gain and travel cost.

Use absence as evidence

When an object is not visible at an inspected site, probability mass is<br>redistributed over the remaining hypotheses and the search order changes online.

Respect dependencies

Actions become eligible only after their graph dependencies are complete.<br>The teabag cannot be inserted before the mug is placed, and pouring waits for<br>both hot water and the teabag.

Update state beliefs

Inspecting, filling, boiling, pouring and steeping alter beliefs about hidden<br>states. The task finishes only when the model is sufficiently confident that<br>the tea has been brewed and served.

Search→<br>Locate→<br>Move objects→<br>Inspect kettle→<br>Fill & boil→<br>Pour & steep→<br>Serve

Foundational demonstration

Belief-guided search in a realistic 3D home

The original 32-second demo isolates the core inference loop: begin with a<br>semantic prior, inspect useful locations, treat absence as evidence and replan<br>until the cup is found.

Your browser does not support embedded video.<br>Download the MP4 instead.

The video has no audio. All information is presented visually through the<br>first-person environment, posterior belief bars, candidate policy scores and<br>the final cup-detection cue.

Download video

00:00<br>Semantic prior<br>Kitchen 45%, dining table 30%, console 20%, side table 5%.

00:06<br>First inspection<br>The agent tests the most plausible and informative location.

00:15<br>Negative evidence<br>Failure to find the cup redistributes belief over the remaining sites.

00:30<br>Cup found<br>Visual evidence confirms the console and collapses the posterior.

The scientific idea

A world model that can expose its reasoning

The central object is not a fixed action policy. It is a structured belief about<br>hidden states of the world, together with a...