M.Sc. Mobile Robotics · Physical AI · Robot Learning · Simulation · 3D Vision · C++ / Python
Building reproducible robotics and perception systems: high-throughput simulation, sim-to-sim transfer, 6D pose evaluation, video object detection, sensor fusion, and safety.
Measured results and verified behaviors are linked to reproducible commands and committed artifacts in each repository.
| Project | What was measured or verified | Result |
|---|---|---|
| C++ / MuJoCo Sim-to-Sim | Same trained PPO policy · 100 held-out seeds · randomized domain · C++ kinematics vs. MuJoCo rigid-body contact | Randomized: 43% → 55% success · 32% → 8% collision |
| ROS 2–Unity Digital Twin | Live Gazebo → ROS 2 → Unity telemetry and deterministic safety watchdog | 28–30 Hz odom · 20 Hz safety status · ≤0.55 s stop bound |
| LLM → ROS 2 Agent | Natural-language goal → schema validation → safety clamp → deterministic ROS 2 execution | Tool allowlist · turtlesim + Gazebo · JSONL audit trail |
| BOP 6-DoF Pose Evaluation | Deterministic LM-O RGB-D subset · FPFH + RANSAC + ICP · 391 instances | 0.3657 ADD(-S) recall · 0.5314 AUC · 467.90 ms |
| BuzzSet YOLOV++ | Two-stage temporal detector on the four-class BuzzSet validation set | 7.8% → 26.8% mAP · 47.7% AP50 |
| ROS 2 C++ Safety Node | C++ safety-filter callback · 10 M iterations · -O2 |
40 ns · 24 M calls/sec · 5 guard layers |
|
C++ / MuJoCo Sim-to-Sim Transfer The same seeded navigation task runs in a fast C++ kinematic backend and MuJoCo rigid-body simulation. A checked-in 100-seed transfer run exposes the success and collision gap directly.
|
Safety-Aware ROS 2–Unity Digital Twin Gazebo drives the TurtleBot3 while Unity mirrors pose and lidar. A deterministic supervisor demonstrates obstacle stops, frozen-scan handling, and operator e-stop behavior.
|
|
A tool-calling agent translates operator intent into allowlisted, schema-validated, safety-clamped ROS 2 actions across turtlesim and Gazebo backends.
|
Reproducible LM-O RGB-D evaluation with native ADD/ADD-S metrics, a classical FPFH–RANSAC–ICP baseline, failure accounting, and committed reports.
|
|
BuzzSet Temporal Video Object Detection YOLOV++ temporal aggregation and YOLOX-Swin experiments improved four-class pollinator detection while exposing class-level failure patterns.
|
ROS 2 C++ Teleoperation Safety Node Five independent C++ guard layers—velocity clamp, acceleration limit, deadman timeout, e-stop, and obstacle stop—running live in Gazebo Sim.
|
These are the strongest end-to-end systems and measured benchmarks; the
mobile-robotics-*algorithm studies are grouped farther down.
| Project | Highlights | |
|---|---|---|
| 🔵 | C++ / MuJoCo Navigation | C++17 + pybind11 + Gymnasium · PPO · 860k steps/s native batch · domain randomization · sim-to-sim transfer |
| 🔵 | Safety-Aware ROS 2–Unity Digital Twin | ROS 2 + Unity + Gazebo · C# · lidar fault injection · deterministic safety supervision · 16/16 tests |
| 🔵 | LLM → ROS 2 Command Interface | Tool-calling · safety clamp · schema validation · JSONL logging · turtlesim + Gazebo |
| 🔵 | BOP 6-DoF Pose Benchmark | BOP/LM-O RGB-D harness · ADD/ADD-S · FPFH + RANSAC + ICP · deterministic reports · Docker + CI |
| 🔵 | BuzzSet YOLOV++ | 4-class pollinator video detection · YOLOV++ temporal pipeline · YOLOX-Swin · RF-DETR comparison · COCO AP |
| 🔵 | ROS 2 C++ Teleoperation Safety Node | 5 guard layers · 40 ns callback · CMake · CI · Gazebo Sim · 11 YAML params |
| 🔵 | NavViz Unreal | Unreal Engine 5 C++ renderer · PPO world-state streaming over TCP/JSON · real-time 3D visualization |
| 🔵 | GNSS-Denied Visual-Inertial Localization | EKF sensor fusion · Docker · ROS 2 C++ wrapper · regression tests · 0.24 m outage RMSE |
| 🔵 | Explainable Vision Demo | ResNet · EfficientNet · ViT · GradCAM · attention rollout · 3 Streamlit apps · CI |
ROS 2 & Systems
- ROS 2 Multi-Robot Coordination — coordinator/agent nodes, namespaced topics, waypoint assignment, Gazebo Sim
- DJI ROS 2 Aerial Mapping Pipeline — mission manifest intake, GeoJSON footprint export, ROS 2 handoff
- Teleoperation Safety Interface — browser HMI, safety interlocks, scenario presets, GitHub Pages
- Robotics Edge Infra Lab — Ansible, Jetson provisioning, Jenkins CI, Docker for edge nodes
ML & Perception
- Machining Digital Twin Microservice — FastAPI, Docker, live dashboard, Swagger UI
- Machining Vibration ML — physics-guided Random Forest for tool wear classification from vibration signals
- Surface Roughness Signal Prediction — RF regression on machining acoustic features, public dataset ready
- UAV Building Footprint Extraction — k-means + MRF segmentation, polygon vectorisation from aerial imagery
Probabilistic Robotics
- FastSLAM · EKF Localization · Monte Carlo Localization · Occupancy Grid Mapping · Bayes Filter · Odometry Motion Model · Landmark Observation Model
Planning & Classical CV
- C++ Grid Path Planner · MDP Value Iteration · Grid Path Planning A*
- Camera Calibration + Pose Demo · Panorama Stitching · Classical Image Processing
- Graph Cut Segmentation · Hough & Mean Shift · MOG People Tracking · Kalman Filter Tracking · Statistical Shape Modeling
Simulation & Physical AI: MuJoCo, Gymnasium, Gazebo, Unity, Unreal Engine, PPO, domain randomization, sim-to-sim transfer, reduced-order modeling
Robotics: ROS 2, rclcpp, rclpy, colcon, nav_msgs, sensor_msgs, tf2
Languages: Python, C++17, C#, JavaScript, MATLAB, Bash
ML / Vision: PyTorch · OpenCV · scikit-learn · Streamlit · GradCAM · SLAM
Infra: Docker · Ansible · Jenkins · GitHub Actions · FastAPI
Hardware: Arduino · PCB design · PLC/SCADA · embedded sensors and motor control
- Mobile robot autonomy and field-deployable perception systems
- Safety-aware teleoperation and human-supervised robot control
- Sensor fusion and GNSS-denied localization
- LLM-to-robot command interfaces
- Multi-robot coordination
Open to research assistant, thesis, and robotics engineering opportunities in Physical AI, robot learning, simulation, perception, and safety-aware autonomy.
