Agents that act from beliefs
Designing controllers that infer hidden state, evaluate possible futures and act while tracking uncertainty, rather than simply mapping observations to actions.
Consulting, advisory and research collaboration
Uncertainty-aware world models for physical AI, robotics and adaptive decision-making.
I build and advise on systems that maintain beliefs about hidden states, act under uncertainty and recover when observations do not fit predictions. The work combines active inference, generative models, dynamical systems and neurocomputational modelling.
Active inference
World models
Embodied AI
Robotics
Bayesian inference
Adaptive decision systems
What this is
A practical route from neuroscience-inspired theory to inspectable AI systems.
I am especially interested in systems where behaviour depends on incomplete evidence: robots in cluttered spaces, autonomous agents that need to search and revise beliefs, decision systems with changing context and scientific data where the hidden causes matter more than surface prediction.
Internal models that can be inspected
Building small but explicit generative models of objects, scenes, observations, policies and failures, so behaviour can be debugged mechanistically.
Planning beyond a fixed reward
Using active inference and expected-free-energy-style objectives to balance goal pursuit, safety, information gain, stability and changing context.
Applied case studies
See how the research maps onto real autonomy problems.
Two short commercial case studies translate the tea-making world model and the drone controller into the kinds of problems they address: autonomous task execution under incomplete information, and embodied control when perception is uncertain.
New embodied world-model demo
From finding an object to completing a household task.
The newer Habitat-Sim demonstration extends the cup-search world model into a full tea-making task. The agent must decide what to search for, revise its beliefs as evidence arrives and coordinate a sequence of actions whose success depends on hidden object states.
Search, infer, act and verify.
The task is represented as a dependency-aware sequence rather than a fixed animation. At each stage, the agent combines its current beliefs, the relevance of each object to the task, expected information gain and movement cost before selecting what to do next.
Active search
Candidate object-location searches are compared online, allowing useful discoveries to be made even before they become the immediate task bottleneck.
Negative evidence
Not seeing an object where it was expected redistributes probability across the remaining locations and changes the next search decision.
Stateful task execution
The agent checks whether the kettle contains water, fills and boils it when required, then inserts the teabag, pours, steeps and serves.
More research prototypes
Further demonstrations behind the applied work.
These prototypes show the wider research programme behind the case studies: belief-based control, partial observability, uncertainty-sensitive planning and multi-objective decision-making across robots, gridworlds and real-time agents.
Drone control
View demo and write-up →
Polyphonic active inference drone
A 3D PyBullet drone controller with noisy self-observations, egocentric target cues, ray-based obstacle sensing, belief-state control and scene-aware policy evaluation.
World model
Open world model lab →
House-search semantic world model
An embodied agent searches for a cup using semantic priors, planned inspection, negative evidence, belief updating and route redirection.
Embodied AI
View robot page →
Active inference robot
A 3D physics example where perception, action, correction and recovery are tied together through a generative model and active inference loop.
Multi-objective control
View gridworld write-up →
Polyphonic gridworld
A partially observed gridworld where safety, goal pursuit, energy maintenance, uncertainty reduction and habit are kept explicit and negotiated online.
Real-time inference
Open Pong demo →
Thermodynamic VL Pong agent
A compact real-time example of posterior updating and action selection through a lightweight thermodynamic variational Laplace scheme.
Research programme
Read the public record →
Coordination without command
A broader research direction on how coherent behaviour can emerge without a single central controller, linking active inference, robotics and adaptive intelligence.
Why it matters
The point is not just another agent demo.
Many AI systems behave impressively until the world changes, evidence becomes partial, or the task requires them to know when they do not know. My interest is in agents and decision systems with explicit internal state: what they believe, what they expect, what they are uncertain about and why a particular action was chosen.
Consulting and advisory
Clear ways to work together.
Best fit: difficult early-stage R&D problems where a principled model, architecture, simulation or proof-of-concept can reduce technical uncertainty before a larger build.
Technical diagnostic / architecture review
A focused engagement to understand the system, identify where uncertainty or hidden state is causing difficulty, and define a technically credible route forward.
From £1,500
- Problem and architecture review
- Focused technical working session
- Written recommendations, risks and prototype route
Prototype sprint
A tightly scoped modelling or simulation sprint that turns the core idea into something inspectable, testable and useful for the next technical decision.
£7,500–£15,000
- Defined question, baseline and milestone
- Working simulation or modelling prototype
- Code, results and technical interpretation
Applied autonomy / modelling PoC
A larger proof-of-concept for teams testing a new autonomy, world-model or uncertainty-aware decision architecture against a concrete use case.
£15,000–£30,000+
- Problem-specific architecture and implementation
- Comparative experiments and failure analysis
- Technical handover and next-stage recommendations
Ongoing / fractional advisory
For teams that need recurring scientific or technical input rather than a fixed build: architecture review, experimental design, model critique, technical strategy and selected hands-on R&D support. Scoped around a defined level of involvement.
By arrangement
Also open to
Fractional research roles, technical advisory positions, invited talks, research partnerships, academic-industry collaborations and selective work in neurotechnology and mechanistic modelling.
Contact
Bring a hard problem, a possible collaboration, or a half-formed idea.
For consulting, advisory work, speaking or research partnerships, email me with a short description of the problem and what kind of help would be useful.
Suggested subject line: Consulting / world models enquiry