University of Essex Human-Centred AI Lab Centre for Computational Intelligence

Research

Three pillars, one laboratory.

The Lab studies AI that acts in the physical world, AI that can be audited, and AI that is built to work with people. The themes below are the ones the group actually works on.

Much applied AI stops at a model that scores well on a held-out set. The Lab’s questions start after that. Can the system perceive a messy physical scene and do something safe with what it sees? Can a person tell why it acted, and when not to trust it? Does it know when the decision belongs back with a human being?

The Lab is part of the Centre for Computational Intelligence. The work is carried by the Lab and by two collaborators at the University of Essex. The same material is taught: see Education.

Pillar 1

Physical AI

From seeing to doing

A depth camera on a desk tripod
Depth camera. Marc Auledas, CC BY-SA 4.0. Cropped.

Physical AI is the step from a model that labels an image to a machine that can act. The Lab works on collaborative and autonomous robots, on large vision models, and on fusing camera, LiDAR, depth, thermal and event sensing into a 3D picture that is current enough to use.

Where a round trip to a data centre is too slow, the model has to run on the hardware in the room: edge devices, NPUs and FPGAs. Prediction is not the end of the pipeline. A useful system adjusts the operation so that the failure it can see coming does not happen.

Themes

  • Vision–language–action models — robot policies driven by foundation models rather than a script for every case.
  • 3D perception and sensor fusion — LiDAR, depth, thermal and event cameras; SLAM and pose estimation.
  • Sim-to-real reinforcement learning — train where it is safe to fail, deploy where the work is.
  • Collaborative-robot safety — force control, speed-and-separation monitoring, digital twins.
  • Edge AI engineering — quantisation, pruning and distillation, then deployment to NPU and FPGA.
  • Prescriptive operations — from forecasting a failure to correcting for it.
  • Autonomous mobile robots — fleet navigation, docking and safe movement in shared spaces.

Pillar 2

AI alignment and explainable AI

From black box to glass box

Dies across a semiconductor wafer
Semiconductor wafer. DrHughManning, CC BY-SA 4.0. Cropped.

Alignment, for this Lab, is a technical problem as well as a governance one. An explainable model is the engine: it can say why it acted, and it can say how sure it is. Without that, a person cannot approve, override or audit the decision in any way that would hold up later.

The group works on interpretable and neuro-symbolic models, on conformal prediction and calibration, and on privacy and provenance. That includes telling an authentic recording or sensor trace from a synthetic one. Dr Javier Fumanal Idocin maintains Ex-Fuzzy, an open-source library for fuzzy rule-based explanation.

Themes

  • Mechanistic interpretability — attribution, concept probing and circuit analysis for deep models.
  • Fuzzy rule-based and neuro-symbolic AI — systems that state their reasoning in human terms.
  • Conformal prediction and calibration — error rates a user can rely on, and a model that knows when to abstain.
  • Privacy-preserving AI — federated learning, differential privacy, encryption and zero-watermarking.
  • Provenance and media forensics — authentic signals and recordings distinguished from synthetic ones.
  • Assurance and governance — auditing, red-teaming and documentation fit for a regulated setting.
  • Human oversight by design — approval, override and escalation that still make sense in an audit.

Pillar 3

Human–AI teaming

From tools to teammates

A participant in an EEG cap watches a task while neural signals are recorded
Brain–computer interface experiment. Laurens R. Krol, CC BY 4.0. Cropped.

A teammate is not a dashboard. The Lab studies the state of the person doing the work: fatigue, cognitive load and attention, read from the eye and from neurophysiological signals, in settings where a missed cue is a safety problem.

The same line of work covers gaze on ordinary phones and laptops, embedded brain–computer interfaces, and assistive and rehabilitation robotics. Exoskeletons are part of that picture, because they change who can do skilled physical work, and for how long. The systems are designed to hand the decision back when confidence, or the person’s state, says they should.

Themes

  • Brain–computer interfaces — EEG decoding, motor imagery, and embedded BCI on FPGA.
  • Eye tracking and gaze interaction — calibration-light tracking on everyday mobile devices.
  • Operator-state estimation — fatigue, cognitive load and attention in safety-critical work.
  • Wearable physiological sensing — signal processing and artefact rejection outside the lab bench.
  • Assistive and rehabilitation robotics — exoskeletons, neurorehabilitation, inclusive workplaces.
  • Adaptive automation — the hand-over between machine and person, designed rather than hoped for.
  • Inclusive interfaces — hands-free and low-literacy interaction that widens who can do the job.

Across the pillars

Foundation models and the stack they sit on

The three pillars all depend on a layer that is easy to treat as someone else’s problem: how a model is adapted, grounded, monitored and paid for in energy. The Lab treats that layer as part of the research.

  • Language models in production — fine-tuning, LoRA, distillation and evaluation that reports the failures.
  • Retrieval-augmented generation and agents — assistants grounded in a real engineering workflow.
  • Multimodal and time-series foundation models — pretraining on sensor and process data.
  • Self-supervised and stream learning — value from unlabelled operational data.
  • MLOps at the edge — versioning, monitoring, drift detection and a rollback that is safe to use.
  • Green AI — energy- and carbon-aware design and deployment.
  • Synthetic data and simulation — training where real data is scarce, sensitive, or unsafe to collect.

Where it lands

Sectors the work is built for

Semiconductors and electronics

Explainable checks on design verification and wafer-level inspection.

Advanced manufacturing

Cobots, line stops a person can justify, and prescriptive maintenance.

Digital services

Edge deployment, and teams who can govern the systems they ship.

Energy and green tech

Predictive and prescriptive control of energy assets.

Medical devices and health tech

Signal models with privacy built in, at a standard a device maker can defend.

Technicians in clean-room clothing working on precision equipment
Clean-room work, used here only as illustration. Aileen Devlin / Jefferson Lab, public domain. Cropped.

From the lab to a site

A pilot the Lab proposes

One concrete shape for the three pillars together is an open physical-AI cell, hosted on a partner’s own site. Vision and LiDAR inspect a process on edge hardware. The cell can flag a defect and halt the line. It also explains why, with a confidence score an engineer can audit, and it returns the decision to a person when that confidence drops.

It is proposed as hands-on work, so the partner’s trainers leave able to teach it. Scope the site and the use case, build the edge model and the explanation, train a cohort, then scale. It is an offer of joint work, not a cell the Lab claims to have already installed.

How teaching and joint projects are packaged.