RELBot Autonomous Mobile Robot Tracking

Real-Time Object Tracking with Mobile Robot

Project Overview

The RELBot is a differential-drive mobile robot built to track and follow objects autonomously in real time. The project brings together single-camera vision and deterministic motor control on an embedded Raspberry Pi 4, balancing fast on-board processing with remote GPU computation.

Raspberry Pi 4 Xenomai 4 / EVL RTOS Sensor Telemetry Computer Vision & Depth

The Challenge

Making a compact mobile robot follow moving targets smoothly in real time came with several key engineering challenges:

  • Limited Processing Resources & Real-Time Execution: Running vision algorithms while maintaining strict real-time control on an embedded Raspberry Pi 4 required careful resource budgeting, CPU core isolation, and low-level code optimization to prevent computation bottlenecks.
  • OS Timing Jitter: Standard Linux kernels introduce 10 to 15 ms of unpredictable scheduling delays. For closed-loop motor control, this jitter causes jerky movement and unstable tracking.
  • Single-Camera Perception: With only a lightweight monocular camera on board, the robot had to detect targets and estimate distances in real time without the extra weight or power draw of bulky sensor suites.

Engineering Solution & Live Implementation

To solve this, I designed a multi-tier cyber-physical setup dividing tasks between deterministic on-board motor control and flexible vision processing:

  • Firm Real-Time Motor Control: Modeled the robot kinematics and wrote a custom PID controller in low-level C++. Running on an isolated CPU core with the Xenomai 4 / EVL real-time framework, the control loop hits a rock-solid 1 kHz rate with under 15 microseconds of jitter.
  • Hybrid Vision Pipeline:
    • On-Board Edge Processing: Fast morphological image filters run directly on the Raspberry Pi CPU for high-rate line and visual pattern tracking.
    • Workstation GPU Telemetry: For complex target detection, video frames stream over low-latency telemetry to a workstation running YOLOv8 and MiDaS monocular depth estimation, which computes bounding boxes and depth maps before sending velocity commands back to the robot in real time.
  • Live Autonomous Following Tests: In physical trials, the RELBot reliably locked onto a moving engineer wearing a safety helmet and smoothly adjusted its drive motors to maintain an exact 3-meter following distance throughout the run.

Real-Time Object Following Demonstration

Performance Results

The multi-tier architecture delivers deterministic control and stable distance tracking:

< 15 µs
Control Loop Jitter
Deterministic 1 ms cycle on Xenomai 4 / EVL isolated core
1 kHz
Loop Frequency
Low-level C++ PID controller executing on dedicated CPU core
3.0 m
Following Distance
Maintained steady distance with monocular depth feedback
Hybrid
Edge + GPU Telemetry
On-board morphological vision paired with remote deep learning

Technologies

Raspberry Pi 4 Xenomai 4 / EVL RTOS Low-Level C++ Programming PID Controller Morphological Image Processing YOLOv8 MiDaS Depth Estimation Sensor Telemetry ROS2 Humble Monocular Vision