Building robots that
bridge imagination & reality

Overview

A fully gesture-driven system to control 3 simulated DJI Mavic2Pro drones in Webots using only hand gestures captured by a laptop webcam. Two separate neural networks — both built entirely from scratch in NumPy — classify left-hand pose (which drone to select) and right-hand pose (what action to perform), with results written to a shared JSON file that each drone's controller reads every 8ms.

Key Achievements

• Two-hand control scheme: Left hand selects drone(s) — 1/2/3 fingers or open palm for all. Right hand commands action — Takeoff, Land, Altitude Up/Down, Flip Forward, Hover
• Custom 5-layer FC-NN from scratch: 35→64(ReLU)→32(ReLU)→16(Tanh)→N(Softmax), with L2 regularisation and cross-entropy loss — no PyTorch, no TensorFlow
• Three optimisers implemented: Adam (first-order), L-BFGS (quasi-Newton, second-order), Nelder-Mead (zero-order gradient-free) — compared experimentally with 5-fold CV
• 35-feature hand vector: 10 fingertip relative positions, 5 extension angles, 3 wrist orientation values, 2 palm centre offsets, extracted per hand from MediaPipe HandLandmarker (21 landmarks)
• PID altitude control + motor mixing: P-controller (K_P=3.0) drives 4 motors per drone; mass compensation adjusts thrust proportionally when payload changes mid-flight
• Flip state machine: 3-stage (Windup 160ms → Rotate 320ms → Recover 240ms) for safe mid-air forward flips at >1.5m altitude
• Fail-safe: No gesture detected for >2s → drones hold position autonomously
• Personally recorded 800–1,200 hand-gesture training samples across 10 gesture classes (4 left-hand, 6 right-hand)

Overview

Real-time multi-object detection and tracking on Raspberry Pi 4 using YOLOv8 (NCNN-exported) with two MG996R servos for pan-tilt camera control.

Key Achievements

• Picamera2 + YOLOv8n NCNN model for low-latency inference on Pi hardware
• PID-style proportional control keeps selected class centred in frame
• Configurable target class (default: person), deadband, max step angle, gain
• External 5V servo power supply; pigpio for PWM on GPIO17/18

Overview

Dynamic waypoint scheduling with occupancy detection for autonomous mobile robot navigation using Nav2 and SLAM on TurtleBot 4. Deployed in a real indoor environment (H101 lab map at Coventry University).

Key Achievements

• Built dynamic_occupancy.py — server tracking FREE/OCCUPIED state per waypoint and queue slot in real time
• Built interactive_navigator.py — client sending Nav2 goals, auto-rerouting to queues when waypoints are occupied
• 6-waypoint system including pre-dock position and full autonomous docking sequence
• Deployed on TurtleBot4 with ROS2 Jazzy + AMCL localization on Ubuntu 24.04
• 5-terminal launch sequence: localization, Nav2, RViz2, occupancy server, interactive navigator

Overview

A dual-Arduino autonomous robot that maps a room using HC-SR04 ultrasonic sensors and encoder odometry, then executes a planned return journey with 5 equally-spaced scheduled stops. Clean Master/Slave architecture separates sensing from motor control across two Arduino Uno boards.

Key Achievements

• Master (sensor board) sends commands <MAP>, <RUN>, <STOP>, <SLOW> to the Slave (motor board) via serial
• Encoder-based odometry: 850 counts = 1 wheel revolution — minimum detectable movement ~7μm
• Bridge detection via upward-facing HC-SR04: auto-slows to 50% speed through narrow passages, then resumes
• Rear obstacle safety: unexpected obstacle triggers emergency stop with red LED indicator
• DRV8833 motor driver with tunable LEFT_SPEED_CORRECTION constant for drift compensation
• Configurable number of stops via NUM_INTERVALS constant — no re-wiring needed

Overview

A fully autonomous three-component lab robotics system that combines mobile navigation, 6-DOF manipulation, and RFID identification to sort test tubes into the correct bins without any human intervention. TurtleBot4 transports loaded tube holders from a pickup point to the docking station, triggering the Niryo NED2 arm via HTTP REST API to pick each tube, scan it with an ESP32+PN532 RFID reader, and place it in the correct bin slot — all orchestrated through event-driven threading.

Key Achievements

• Three-layer architecture: Hardware (NED2 + TurtleBot4 + ESP32), Software (ROS2 + PyNiryo + Flask), Integration (HTTP REST API + threading events + thread-safe state)
• Flask REST API with 6 endpoints: /start_niryo, /rfid, /niryo_done, /status, /emergency_stop, /reset — enabling fully decoupled multi-robot coordination
• Event-driven RFID sync: threading.Event() blocks the Niryo thread at the scanner position until the ESP32 sends the tag UID via WiFi HTTP GET, then unblocks to route to the correct bin
• 4-bin Cartesian sorting: Niryo uses hardcoded joint-space waypoints (Approach → Place Approach → Place) at SLOW_SPEED (20%) for safe tube insertion; 4 UID-to-bin mappings in rfid_bin_map
• ESP32 + PN532 RFID scanner: I2C on GPIO21/22, MIFARE ISO14443A, <5cm range — reads tag UID and fires HTTP GET to Flask server; Arduino C firmware with WiFi, Wire, HTTPClient libraries
• TurtleBot4 coordinator: ROS2 Nav2-driven loop — navigate to pickup → dock → trigger Niryo → poll /niryo_done → undock → clear costmaps → repeat
• Collision recovery: Auto-clears collision detected flag and retries movement; safe fallback to home position on critical errors; emergency stop via learning mode enable
• 4 test tubes per cycle across 2 bins (A & B) with 2 slots each; post-dock costmap clearing + 2s AMCL settle prevents Nav2 localisation drift between cycles

Overview

This slot is reserved for upcoming projects currently in development. Visit my GitHub profile to see the latest work in progress.

Skills & Tools

Context

Python primary for robotics. C++ for performance-critical nodes. Bash for system scripting.

Skills & Tools

Context

Full ROS2 Jazzy systems on Ubuntu 24.04 — Nav2, SLAM, MoveIt2 for real robot deployment.

Skills & Tools

Context

Custom 5-layer NNs from scratch. MediaPipe HandLandmarker. OpenCV ArUco pose estimation. YOLOv8 NCNN for RPi.

Skills & Tools

Context

Hands-on with real robots. Debugged I2C, WiFi, TCP, serial comm. Solved mechanical offset issues.

Skills & Tools

Context

Webots and Gazebo for simulation. Flask REST APIs for multi-system coordination. All on GitHub.

Skills & Tools

Context

Applied thermodynamics, exergy analysis. Familiar with ISO 31000 risk and ISO 10218 robot safety.