Curriculum vitae
Robotics, systems and control engineer. Mechatronics and physics. Sydney; Australian citizen (E-3 visa eligible). Available from January 2027.
Now
Oct – Dec 2026: full-time research sabbatical; available January 2027. I work on both the hardware and the theory: a 30,000-line C++ stack that flies a 10-drone flock, a braid-group safety certificate for tethered drone swarms with its own frictional-contact solver, and certified bounds on controlling organised systems. On sabbatical I'm finishing three papers, building a GPU physics engine, and researching and writing about AI, physics and robotics.
Research
UAV flock herding (Ferroflock)
- Built a 30,000-line C++ autonomy stack for a 10-drone Tello/OptiTrack testbed and flew it as a flock, herded live in the lab (control walkthrough, all videos). A 50 Hz Kalman filter keeps position through tracking dropouts, and setup tooling cut lab setup from half a day to ten minutes.
- Knowing the flock's current shape gives most of the herding benefit; predicting individual birds only pays off for a single fast drone. Herding from the flock's edge keeps it together, which a rigid-body flock model can't capture.
Tethered drone swarms: a braid-group safety certificate
- Proving that drones tethered to a shared collar can move without tangling, with a safety certificate built on the braid group and control barrier functions keeping the swarm in a forward-invariant safe set.
- In a discrete-elastic-rod simulator, fixed endpoints don't determine a crossing's state once it slides (0.53 µm vs 5.9 mm slip under the same probe), but the hitch position does. Found three silent defects in the DisMech solver and submitted fixes upstream; the first is merged.
- Built a frictional-contact solver for the certificate that reproduces the thick-cable capstan law, where MuJoCo's default contact holds a ratio of only 1.01–1.14 against an ideal of 6.6. Now porting it to a batched CUDA engine benchmarked against MuJoCo.
Identifiability in automated drilling
- Drilling planners that chase lower predicted uncertainty can stop collecting the close sample pairs they need, and nothing in their scores flags it. From sparse starts, 54% of first batches left the Gaussian-process range unidentified. Traced it to vanishing Fisher information, built a per-batch guard that fixes it without losing accuracy, and checked it against 2,939 Western Australian drilling campaigns.
The cost of keeping a system organised
- A system can hold its structure up to a deadline at no cost yet be unable to hold it afterwards, so requiring it to stay organised past the deadline costs real performance. Interval-certified examples in a 3D bird flock, a power-grid generator and a superconducting condensate.
AI research
- Is slop load-bearing? Testing whether the stylistic habits of LLM output carry calibration and honesty or are vestigial, by suppressing fourteen families of AI-isms and measuring calibration, abstention and sycophancy. Grew out of Tacita's agent systems and an essay series (first essay published).
- Tally: do coding agents understand, or recall? The test is a made-up language that looks exactly like Python but works differently in nine places. Frontier agents got through; a smaller open model kept writing Python.
Camera simulation for long-range drone detection
- Physics-based synthetic data to predict how a trained detector performs on cameras it was never tuned on, so a camera can be designed before it's built. In early tests on an unseen camera, simulation-trained detectors found far more drones than ones trained on real footage from another camera.
Also: Superconductivity: for picosecond current pulses, the safe drive is set by the gap's minimum after the pulse, not at its end; in a certified model, the end-of-pulse view overstates it by at least 20%. Wind waves: a first-principles model of how wind feeds ocean waves, with no tuned constants, matching 25 published simulation growth rates within a median factor of 1.2; GPU spectral-element simulations (NekRS) to test it.
Experience
Tacita Pty Ltd
- Legal AI. An on-premise platform over 500k+ documents. Delivered a 5,000-email eDiscovery matter end-to-end in under a week, replacing months and A$30–40k of manual work (lawyer estimate), with staged adversarial verification of the outputs.
- Ingestion at scale. Indexing had silently slowed from 800 to 20–30 documents a minute while the pipeline reported success, putting the database load at months. Fixed the root causes and gated completion on store-versus-source counts; a 200k-document load now takes an afternoon.
- Also built Substrate, an agent-native systems engineering platform for long-horizon human and agent work; Episteme, a Bayesian belief engine that treats sources like sensors in a state estimator; and Altus, property feasibility with correlated Monte Carlo.
Raytheon Australia
- Automated Australian Defence Air Traffic System engineering workflows, saving 500+ hours a year across a five-person team. Systems architecture in DOORS, Windchill and Cameo; held baseline security and export clearances.
Rio Tinto Centre for Mining Automation
- Adaptive sampling on Pilbara drill-core data, cutting sampling density 30% in sandbox experiments with up to 50% more information per sample than uniform sampling.
University of Sydney
- Tutored Sensors and Signals (2024). Australian Centre for Field Robotics Research Scholar on neural underwater change detection (2022–23). Physics Special Studies Program research projects (by invitation): piezoelectric nanofibres (2021), and MOF-glass perovskite solar cells with Anita Ho-Baillie's group (2023), including device fabrication and characterisation.
Publications and manuscripts
- F. Wilson, A. Melkumyan, A. J. Hill. Variance-Reduction Infill Sampling Can Leave the Variogram Unidentifiable: A Per-Batch Diagnostic and Guard for Automated Drilling. For IEEE T-ASE; preprint soon.
- F. Wilson, Z. Wang. Performance Bounds and the Cost of Continuation under State Constraints. Preprint soon.
- F. Wilson, Z. Wang. When Does Flock Prediction Help UAV Herding? Shape, Speed, and Execution Limits. For IEEE RA-L; in preparation.
- F. Wilson (2024). The value of information in ML-based mine geology modelling. Honours thesis, First Class.
Education
University of Sydney
- BE (Honours Class I), Mechatronics/Space Engineering; BSc, Advanced Physics.
- Selected marks: Aerospace Engineering Project 3 (95), Mechatronics 3 (92), Physics Interdisciplinary Project (91).
- Projects: SLOSHSat CubeSat fluid-slosh payload, where I designed the custom PCBs for sensing, localisation, experiment control and magnetorquer actuation, wrote the firmware on FreeRTOS, and 3D-printed the structure in carbon-fibre polycarbonate (PC-CF); quadrotor autonomous landing on a moving ground vehicle; TurtleBot SLAM.
- Leadership and awards: led the 25-person Sydney University Rover Team through critical design review; executive of the Physics and Engineering societies; Dean's List (2024); Australian Centre for Field Robotics Research Scholar (2022–23); 1st place, AIAA Aerohack; Raytheon GradX, Engineering Sydney Industry Placement and Sydney Innovation Program scholarships; LSAT 174; GRE 334 (169 verbal, 165 quant, 5.5 writing); Optiver Quant Test 80/80.
Technical skills
- Languages and tools
- C, C++, Python, CUDA, Verilog (FPGAs), MATLAB, PyTorch, TensorFlow, ROS, Docker, Linux, Git, SysML
- Hardware
- PCB design, embedded firmware, real-time operating systems, 3D printing and CAD
- Simulation and physics
- Physics simulation, multibody and contact dynamics, numerical methods, sim-to-real
- Control and estimation
- State estimation, Kalman filtering, sensor fusion, model predictive control, optimal control, control barrier functions, Lyapunov analysis, decentralised multi-agent control
- ML and verification
- Gaussian processes, Bayesian inference, computer vision, reinforcement learning, LLM agents and evaluations, interval-arithmetic verification