Skip to content

2024 — Python

Repository

github.com/UM-Driverless/driverless  •  Video tutorial

The original autonomous system, built as a single Python application. All components run in one process orchestrated by main.py.

Architecture

main.py
 ├── YOLOv5 cone detection (PyTorch + CUDA)
 ├── ZED camera interface (pyzed SDK)
 ├── Path planning
 └── CAN bus control (Kvaser canlib)

Pipeline: Camera frame → YOLOv5 inference → depth projection via ZED → path planning → steering/throttle commands over CAN.

Key Components

Component Description
Perception YOLOv5 model trained on Formula Student cones. Weights at src/driverless/yolov5/weights/.
Camera Stereolabs ZED — provides RGB + depth. Configured via ZED SDK.
Control CAN bus communication with actuators via Kvaser interface. Setup scripts: setup_can0.sh, enable_CAN.sh.
Simulator Formula Student Driverless Simulator (FSDS) — AirSim-based, Unreal Engine. Config in src/driverless/config.yaml.

Hardware Targets

  • NVIDIA Jetson Xavier NX — original onboard computer
  • NVIDIA Jetson Orin — upgraded compute platform
  • Desktop with NVIDIA GPU — development with CUDA 12.x / Ubuntu 24.04

Setup

# Python environment (pyenv + venv)
pyenv install 3.12.3
pyenv local 3.12.3
python -m venv .venv
source .venv/bin/activate
pip install -e .

See the repository README for full setup instructions including CUDA, ZED SDK, and Jetson-specific configuration.

Limitations

This architecture served well for initial development but had some drawbacks that motivated the move to ROS 2:

  • Monolithic — all components tightly coupled in one process; a crash in perception takes down control
  • No standard message format — data passed between functions via custom Python objects
  • Hard to test in isolation — can't run perception without the full pipeline
  • Limited simulation — FSDS requires Windows or a powerful Linux desktop with GPU