Stanhope AI raised £2.3 million in seed funding to develop machines that learn from their surroundings using principles drawn from neuroscience. The round was led by the UCL Technology Fund, with participation from Creator Fund, MMC Ventures, Moonfire Ventures, Rockmount Capital and angel investors. The company is a University College London spinout supported by UCL Business.
Its founding team combines neuroscience and applied AI expertise. CEO Professor Rosalyn Moran previously served as deputy director of King’s Institute for Artificial Intelligence. Director Karl Friston is a professor at the UCL Queen Square Institute of Neurology, while technical adviser Dr Biswa Sengupta has worked on AI and cloud products at JPMorgan Chase. Their approach centres on active inference and the free-energy principle.
Active Inference Starts With Prediction
Stanhope’s algorithms continually predict what will happen next and compare those expectations with incoming sensor data. When the observation differs from the prediction, the system updates its internal model of the world. This resembles the company’s account of how brains adapt their view of an environment while trying to reduce surprise and uncertainty.
The model therefore learns during operation rather than relying only on patterns fixed during a large pretraining process. Onboard sensors provide a continuous stream of evidence, allowing an autonomous machine to refine its understanding as conditions change. That capability is particularly relevant to physical systems, because a drone or robot cannot assume that every environment will match its training examples.
The Approach Differs From Large Language Models
Traditional language models learn from vast collections of previously seen data and generate likely responses within that learned structure. Stanhope argues that such systems cannot independently rebuild their understanding while acting in the world. They also require substantial compute and energy for training and operation. Active-inference models are intended to be smaller, adaptive and more directly connected to current sensory evidence.
The company further claims that minimising uncertainty by design can reduce hallucination. This idea should be understood in the context of its specific models: continuously checking prediction against observation gives a physical agent a mechanism for correcting its world model. It does not mean that every uncertain decision disappears. Sensors can fail, assumptions can be wrong and a system still needs testing under unusual conditions.
Compact Models Fit Drones and Autonomous Machines
Lower size and energy requirements allow Stanhope’s models to run on small devices, including drones. Processing on the machine can reduce the delay involved in sending sensor data to a remote server and enable operation where connectivity is unreliable. It can also let the agent respond to a changing environment while keeping its learning loop close to the sensors and controls.
The company was testing the technology with delivery drones and autonomous machines alongside partners including Germany’s Federal Agency for Disruptive Innovation and the Royal Navy. Longer-term applications could include manufacturing, industrial robotics and embodied AI. These are environments where adaptive decisions matter, but where safety and predictable behaviour must be demonstrated before autonomy expands.
Funding Moves Fundamental Research Toward Application
The investment will support further development of Stanhope’s agentic models and their practical use. This transition is difficult for research-heavy startups. Mathematical ideas must be converted into reliable software, integrated with hardware and tested outside controlled demonstrations. The funding gives the team resources to pursue that engineering while maintaining the scientific foundation that distinguishes its approach.
The free-energy principle behind the work was developed by Friston and describes living systems as acting to minimise surprise and uncertainty. Moran specialises in applying active inference to AI, while Sengupta’s research includes dynamical systems, optimisation and energy efficiency. The founding group therefore connects theory with the system-level questions that determine whether a model can operate on constrained machines.
Investor interest also reflects demand for alternatives to ever-larger language models. A compact system that learns from real-time evidence could address tasks where a static cloud model is too slow, power-intensive or detached from the physical environment. Yet the approach must prove that its advantages persist in messy, safety-critical settings and that its decisions can be understood by operators.
Stanhope AI’s proposition is ultimately about a different route to machine intelligence. Instead of scaling stored knowledge alone, it emphasises prediction, perception and continuous correction. The seed round will help test whether decades of neuroscience research can produce commercially useful autonomous systems. Success will depend less on sounding human in a conversation and more on making dependable decisions while the world changes around the machine.
Delivery drones provide a demanding early test because they combine movement, uncertain surroundings and limited onboard power. A useful system must update quickly without consuming the energy needed for flight. It must also recognise when uncertainty remains too high and defer or stop rather than treating every prediction as sufficient evidence for action.
Partnerships with public innovation bodies and the Royal Navy can expose the models to conditions that a laboratory cannot reproduce fully. Those trials should help Stanhope identify where active inference performs well and where additional safeguards or conventional controls are required. Practical feedback may reshape the technology as much as further theoretical work.
The investment gives the founders time to build that evidence. For a science-based startup, credibility comes from connecting a mathematical account with repeatable results on real machines. If compact active-inference models adapt reliably while using less energy, they could complement rather than simply replace other AI approaches, handling the continuous perception and decision loops for which they were specifically designed.
