PhD Studentship: RL-Based Safe Autonomous Systems for Long-Horizon Robotic Tasks

This 3.5-year PhD project is fully funded. Home and EU students are welcome to apply. The successful candidate will receive an annual tax-free stipend set at the UKRI rate (£21,805 for 2026/27) and tuition fees will be paid. We expect the stipend to increase each year. The start date is July 2026.

We recommend that you apply early as the advert may be removed before the deadline.

Avionic control demands dissimilar redundancy to avoid single-point failure. Neural controllers introduce non-determinism and distribution shift, making classical numeric voters unreliable. This project fuses AI redundancy with safe reinforcement learning to create a certifiable, dependable control stack for UAVs (and more widely to robotic systems).

We will develop families of dissimilar safe-RL policies via architectural diversity, randomisation seeds, objective/constraint variations, and training data perturbations. Safety will be embedded through constrained MDPs, shielded RL with control-barrier/Lyapunov constraints, and offline RL for rare-event coverage. Policies will be deployed across heterogeneous hardware (general-purpose CPUs, GPUs, and an embedded AI accelerator) to meet dissimilarity requirements.

Central to the work is a trust-aware comparator/voter that aggregates proposed actions using calibrated uncertainty (e.g., ensemble disagreement, conformal risk bounds) and temporal-logic guards. A runtime assurance layer supervises the AI stack, handing control to a simple, verifiable fallback when confidence drops or constraints are at risk.

Evaluation will proceed in high-fidelity UAV simulation and hardware-in-the-loop, with fault injection (sensor dropouts, latency, actuator saturation) and adverse environments (wind gusts, GPS loss). We will map evidence to certification guidance (e.g., DO-178C/DO-254 context, ARP4754A objectives), delivering: (i) comparator algorithms and proofs/guarantees, (ii) an open testbed and datasets, and (iii) performance/safety benchmarks for redundant AI-in-the-loop flight control.

The project is expected to commence by the end of July 2026. To support this timeline, shortlisted candidates may be invited to interview during June 2026, with a rapid turnaround in selection decisions.

Applicants should hold, or expect to obtain, a first-class or strong upper second-class degree (or equivalent) in computer science, engineering, or a related discipline. Candidates should also have demonstrated experience in reinforcement learning, machine learning, or AI system implementation.

To apply, please contact the main supervisor; Dr Kieran Wood - kieran.wood@manchester.ac.uk. Please include details of your current level of study, academic background and any relevant experience and include a paragraph about your motivation to study this PhD project. Note the study must commence by end of July 2026.

Advert information