MSc Electronic Engineering candidate at Durham University, preparing for PhD research in trustworthy autonomous systems, computer vision, multimodal perception, and experimental sensing.
I work across the boundary between algorithms and physical evidence: defining typed interfaces around learned components, checking whether sensor streams are trustworthy before using them, and packaging experiments so another researcher can reproduce the decision from committed data and code.
Research approach: field evidence → integrity checks → interpretable metrics → explicit safety/quality gates → reproducible reports
| Repository | Research focus | Public evidence |
|---|---|---|
| UAV multisensor diagnostics | Sensor timing, Camera–LiDAR synchronization, trajectory evaluation, failure injection | Importable Python package, CLI, five tests, ATE/RPE and timing reports, CI-reproduced baseline/degraded benchmark |
| Safety-constrained UAV mission interface | Safe boundary between language-level intent and UAV execution | Typed mission contract, JSON Schema, fail-closed policy, threat model, seven tests, deterministic batch evidence |
| Runtime evidence assurance for UAV telemetry | Bounded STPA-informed hazard-to-evidence traceability and persistence-qualified fallback recommendations | Released v0.1.0 synthetic open-loop replay, executable claim/monitor mapping, hash-bound source/data/results, six-job CI |
| Safe neural control certificates | Analytic robust projection around a learned policy for an uncertain sampled-data plant | Released v0.1.0, separate-code-path certificate recomputation, explicit counterexamples, 90 simulated scalar-model closed-loop runs, 36,012 corner/grid checks, six-job CI |
| Decentralized learning stress test | Peer-to-peer learning under non-IID scarcity, malicious updates, synthetic-data poisoning, and membership auditing | Released v0.1.0, pinned UCI Adult source, 54-run CPU benchmark, 19 tests, three verified result chains, Windows/Linux CI |
| Wireless TSN deadline lab | Deadline scheduling across a Qbv-inspired wired boundary and lossy wireless slot abstraction | Released v0.1.0, C++20, GCC/Clang and sanitizer CI, 18 Python evaluator tests, 240 CI runs, 725 hash-verified result files |
| UAV flight-video quality audit | Computer-vision evidence quality and data provenance | Two released field-test videos, SHA-256 manifest, 224 sampled frames, per-frame metrics, SVG timeline, CI reproduction |
| Multichannel thermal validation toolkit | Embedded sensing, defensive protocol parsing, experimental validation | Fictional public protocol, seeded simulator, status-aware analysis, nine tests, synthetic baseline/fault-injected reports |
| Computer vision and autonomous systems portfolio | Integrated research narrative and physical UAV evidence | Stable CV-facing entry point, project map, field media, evidence boundaries, reproducibility instructions |
- Trustworthy multimodal perception: synchronization, provenance, uncertainty-aware diagnostics, and failure analysis for camera/LiDAR systems.
- Safety-constrained embodied AI: structured mission representations and deterministic runtime guards around language- or vision-conditioned components.
- Safe learning-based control: composing learned nominal policies with explicit model-based constraints and falsifiable closed-loop certificates.
- Trustworthy decentralized AI: peer-local learning, non-IID data, malicious-update stress tests, privacy auditing, and explicit protocol boundaries.
- Time-sensitive networked systems: deadline scheduling, clock uncertainty, loss bursts, reliability mechanisms, and provenance-checked simulation.
- Experimental robotics: connecting physical tests with traceable metrics instead of treating a demonstration video as sufficient evidence.
- Embedded and industrial sensing: reliable data acquisition, status-aware decoding, multichannel consistency, and reproducible validation workflows.
Every featured repository separates what is measured from what is simulated. Controlled benchmarks are labelled as synthetic; field videos are presented as physical-test evidence but not as proof of autonomy; commercial identifiers, proprietary protocols, and client data are not published. Each software repository includes tests, versioned configuration, machine-readable outputs, and GitHub Actions.
| Result | Interpretation |
|---|---|
| Multisensor baseline: ATE RMSE 0.0370 m; sync p95 2.206 ms; 0 failed gates | Controlled synthetic baseline for the public diagnostic pipeline—not a field-accuracy claim |
| Multisensor degraded case: ATE RMSE 0.2119 m; sync p95 18.314 ms; 11 failed gates | Fault injection is detected and reported with coded failures |
| Video audit: 224 decoded field-video samples | Demonstrates provenance and image-quality analysis over released real media |
| Thermal baseline: p95 channel spread 0.3026 °C; 0 issues | Seeded synthetic regression fixture—not a product specification |
| Thermal degraded case: p95 channel spread 1.1843 °C; 8 issues | Status faults, channel offset, drift, and spread gates behave as designed |
| Decentralized sign-flip case: peer mean 0.484 vs 0.729 clean synthetic mix; trimmed mean 0.677 under attack | One Adult/all-to-all/three-seed stress test; evidence of a configured failure mode, not a general Byzantine guarantee |
| Wireless release matrix: 240 CI runs; clock-aware EDF improves on FIFO in bursty/retry cases but loses under sync holdover | Retained mixed result from a packet/slot simulator, not a standards, hardware, or universal scheduler claim |
- Email: yuanyuan.tao@durham.ac.uk
- GitHub: @Aleck-Tao