Self-Balancing Robot

Two-wheel self-balancing platform with IMU sensor fusion, PID control, and ROS2 nodes for mapping and person-following.

See Gallery

Overview

Built as a team project: I led system integration and implemented key modules (MQTT (Raspberry PI), computer vision, ROS2 SLAM, Mechanical styling).

Project Goals

Stable balancing via IMU (gyro+accel) with complementary/kalman fusion.
Tunable PID for tilt → motor PWM with safe saturation and anti-windup.
ROS2 nodes: sensor stream, command, SLAM (LiDAR + Cartographer).
Person detection + follow mode (vision or LiDAR tracking).

Role

Team manager: I deployed organisational tool and methods to optimize efficiency and smoothen teamwork: Gantt charts, interative collaborative diagrams, agile strategy.
Computer vision: I implemented haarcascards and Mediapipe computer vision for person following
I integrated ROS2 with a RPLIDA A1M8 to allow live mapping
I contributed to the mechanical design, adding eyes and hat design to make the design stand out

Key Specs

Chassis: Edited from given base and 3D printed
Motors: 12V DC gearmotors + encoders
IMU: MPU-6050 / ICM-20948
Controller: ESP32 / STM32 (PID 1 kHz)
Companion: Raspberry Pi 5 (ROS2)
Sensing: 2D LiDAR (Cartographer)
Power: 3S Li-ion with BMS
Software: C++/Python, ROS2, MQTT

Gallery

Click any photo to view it larger.

Control loop schematic

Role attribution and project team divisions using Miro (Editable by all team members)

SLAM map

Gantt chart

Electronics bay

Force field analysis schematic

Chassis build

Code schematic accross all teams. To manage the project I had to develop an understanding of all subsystems of the robot. This Miro schematic allowed all teams to understand how their contribution interacted with other teams.

IMU mount

Initially, I attemped to use haarcascades lower body model to detect a user's lower body in OpenCV. This was coherent due to the robot's small size, making it see the user's lower body. However, this model appeared to be trained on full body images, making it unreliable when presented with partial bodies.

Full assembly

The most functional approach was to use MediaPipe pose detection with a camera inclined upward. Pose detection appeared more effective at recognizing partial body parts. This was integrated within the base station laptop's UI, and allowed the robot to follow users on demand, by calculating the detection box' offset from center and sending serial commands accordingly.

Chassis build

Gazebo simulation using SLAM toolbox.

IMU mount

Visualization in Rviz2 of SLAM implementation with Google Cartographer

Full assembly

Integration of the camera and mechanical design on the robot.

Demo Videos

Balancing tests, mapping runs, and follow-mode.

Notes - Results

ROS2 Integration

Despite the need to cut some ROS2 features we had wished to implemnt (Nav2), the project was highly successful (1st class) and was the only project implementing a creative design, as seen in the last video.
Comparison was made by a team member of the difference in performance between Mediapipe and Yolo V8. Yolo V8 appeared to be slightly more efficienct but highly more demanding in resources. Mediapipe was implemented. Live UI camera was diabled to prevent power spikes.