Autopilot & FSD - BFMC

The project models a compact ADAS/autonomous driving system: camera, IMU, sonar, GPS/localization, and V2V signals flow through ROS topics; deep learning models estimate lane error, stop-line distance, local path, traffic signs, traffic lights, and obstacles; the controller converts perception into speed and steering commands for the BFMC track. In 2024, the team represented Vietnam and had the opportunity to attend the Romania final, but visa issues prevented physical participation, leaving the team ranked Top 15 worldwide. The project was later open-sourced and expanded as a reference for Vietnamese robotics, academic, and research communities.

A scale-model autonomous vehicle has to run perception, decision-making, and control in real time on very small hardware. The system needed to handle camera streams, lane detection, stop lines, traffic signs/lights, obstacle classification, IMU/sonar feedback, serial communication, and ROS message flow while staying within tight CPU, RAM, thermal, bandwidth, and latency limits.

Team Lead / AI Engineer / Embedded Robotics Engineer

The architecture uses ROS Noetic/catkin with input, control, action, output, perception, utils, and rosserial packages. Raspberry Pi receives images from the RPi Camera V2 over FFC, reads BNO055 through I2C, and communicates with STM Nucleo over UART/USB; the Nucleo drives the steering servomotor, VNH5019 motor driver, DC motor, and AMT103 encoder, powered by a 2-cell LiPo battery through an OKR-T/10 DC/DC converter. Camera nodes publish frames into ROS, the main autonomous loop runs OpenCV DNN with ONNX Runtime-style models, the controller uses Pure Pursuit/PID to produce steering/speed commands, and the serial bridge communicates with MCU hardware for motor and servo control. The AI pipeline is optimized with ROI cropping, small grayscale 16x16/32x32 inputs, small ONNX models, confidence thresholds, tiled-ROI voting, inference timing, and fallback logic to run smoothly on Raspberry Pi and Jetson-class hardware.