Concurrent systems — Java backends, Rust/C++ robotics, video and edge. Dublin.
In structured environments the hardware is catalog. The cost moved to software — perception, planning, control, fleets. I'm proving that software first, in public, without a robot.
| Repo | Focus | What you can check |
|---|---|---|
| Boundaryd | Rust policy gateway | Offline-capable decide API with PostGIS geofences, hash-chained audit, transactional outbox; 2.53M evals/s in-process; cached HTTP 15.3k rps / p50 0.99 ms on a Ryzen 5 3600 |
| Staylock | Java 21 inventory | Hotel reservation ledger with a last-room oracle: 200 concurrent guests → 1 hold / 199 sold out / 0 remaining, asserted in CI; explainable ranking; hold-then-confirm API and public demo UI |
| Repo | Focus | What you can check |
|---|---|---|
| FieldCast | Rust + GStreamer | Real-time two-camera composition and H.264/MP4 output; custom zero-copy continuity element, structured timestamp diagnostics, clean-EOS drain, real media smoke test in CI |
| FrameForge VK | C++20 + Vulkan compute | Chained grayscale/Sobel shaders with explicit barriers and timestamp queries; 1.67 ms at 1080p on RX 5600 XT vs 3.76 ms CPU (2.25x), max error below 3e-7 |
| TrailVision | Multimodal sports video | Auditable voice/IMU/vision event fusion, SQLite replay, FastAPI, and responsive review UI; deterministic demo fuses 17 candidates → 6 events → 3 runs |
| Repo | Layer | What you can check |
|---|---|---|
| arm-lab | Manipulation | Rust UR5e stack; predeclared 300-query simulation study: RRT-Connect 46.0% vs 27.7% direct, velocity-FF 90/138 full passes; independent IK comparison: arm-lab 978/1,000 vs k 718/1,000 with the same eight restart starts |
| synthscan | Perception | Synthetic depth cameras and a point-cloud toolkit in Rust — no PCL |
| girder | Middleware | ROS-free typed pub/sub: 420 ns p50 in-process, bit-exact .gdr replay in CI |
| rust-robotics-toolbox | Foundations | MuJoCo, Bevy, ONNX Runtime, Burn on Vulkan — smoke-tested on AMD |
| robotarium-weighted-coverage | Multi-robot | Fixed-seed eight-robot density-weighted coverage with barrier certificates |
| earth-rover-urban-pilot | Rover readiness | SDK-shaped controller; 100-seed fault matrix cuts commanded outage motion 63.0%; 5,408-frame real-rover-footage replay exercises gross camera-health faults; no hardware run |
Accepted upstream:
| Project | Change | Verification |
|---|---|---|
| ROS Controls #2567 | Adds closed-form bicycle, tricycle, and Ackermann integration tests | Maintainer-approved and merged; exact Rolling package passed 32/32 tests |
Selected patches currently under maintainer review:
| Project | Change | Verification |
|---|---|---|
| Gazebo Sensors #640 | Bounds reused thermal and segmentation message headers instead of appending one entry per frame | Full local suite; Homebrew ARM64, Ubuntu, and Windows hosted builds green |
| ros2-rust #669 | Coalesces duplicate executor task wakes while preserving wake-during-poll | Nine hosted builds green across Humble, Jazzy, Kilted, and Rolling |
| RViz #1837 | Surfaces incompatible QoS as a visible error and recovers after compatible traffic arrives | Fast DDS reproduction; exact core and downstream display checks pass locally |
IROS 2026 Indoor Exploration Competition — CPU-only policy. On a released single-robot map: 65.17% base-station coverage vs 35.02% nearest-frontier baseline in matched local runs. Cross-map and hidden-start robustness testing is in progress.
CS, Trinity College Dublin.