Two compact examples of uncertainty-aware AI in action: a household world model that revises its plan as evidence changes and an embodied drone controller that acts from internal beliefs rather than privileged simulator truth.
A rigid task script assumes that the right object is where expected and that each prerequisite is already known. The tea-making agent instead maintains explicit beliefs about objects and hidden task states, then changes its behaviour as evidence accumulates.
The system must find a mug, kettle and teabag, reason about hidden task state such as whether the kettle contains water and execute a dependency-aware sequence of actions. Its search decisions combine semantic priors, information gain, task relevance and travel cost.
An autonomous agent may know the overall task but not the location or state of the objects needed to complete it. A fixed sequence can fail as soon as one assumption is wrong.
The system maintains independent location beliefs for multiple objects, probabilistic beliefs over hidden task states, a directed task graph and an expected-free-energy-style search policy. Simulator truth remains separate from the agent's internal beliefs.
If an object is not found where expected, probability is redistributed across the remaining candidate locations. If a hidden task state is already satisfied, the corresponding action can be skipped. Evidence therefore changes both search and task execution.
The current implementation is a research prototype, but the architectural pattern is relevant wherever a system must maintain explicit hypotheses, gather missing information and update a multi-step plan.
An embodied system never sees the world from nowhere. Its observations depend on position, motion, occlusion and sensing limits. The drone prototype explores a controller that maintains internal beliefs, predicts possible futures and balances multiple concurrent demands during action.
The controller operates in a 3D PyBullet environment using noisy self-observations, target cues and obstacle sensing. Candidate actions are rolled forward through an internal dynamics model and evaluated against goal, safety, information-seeking, efficiency and memory-related pressures.
A moving platform must continue to regulate itself while observations are noisy, obstacles are only partly observed and task-relevant information may disappear through occlusion or viewpoint change.
The prototype combines a belief state over pose, velocity, occupancy and uncertainty with a learned transition model, short-horizon policy rollouts and a multi-term objective that makes competing control pressures explicit.
Candidate trajectories are not evaluated only by whether they move toward a target. They are also shaped by obstacle risk, uncertain regions, novelty and the need to preserve coherent behaviour as the system's beliefs evolve.
The current work remains simulation-based, but the architecture targets the class of problems where an autonomous platform must reason about uncertainty while remaining continuously engaged with the environment.
Active inference, generative modelling and polyphonic control are technical ingredients. The practical value is in making hidden assumptions explicit, allowing uncertainty to influence behaviour and producing systems whose decisions can be interrogated when something goes wrong.
Keep an internal model of what is likely, what is uncertain and what evidence would change the current hypothesis.
In uncertain settings, the best next action may be the one that improves the system's state estimate before committing further.
Separate beliefs, task state, candidate policies and competing pressures so debugging and counterfactual analysis remain possible.
I work on early-stage R&D, technical architecture, modelling and prototypes for autonomous systems, world models and decision problems where hidden state, partial observability or competing objectives matter.
The strongest fit is a technically difficult problem where a principled model or prototype can de-risk a larger engineering programme.