The effort has moved to DeepRacer Custom!

DeepRacer for Raspberry Pi 4

This repository provides a port of the DeepRacer software stack to the Raspberry Pi 4. It runs on Ubuntu 22.04 with ROS2 Humble, but with most of original code. To make the image smaller the server image is recommended.

Features

Main features of the port:

• Previously trained models will work
• Uses Tensorflow Lite or OpenVINO for inference. Supports the Intel Neural Compute Stick 2 (NCS2/MYRIAD)
• Can integrate with DREM
• Single power source¹ - reduces weight and lowers center of gravity
• Over-the-Air Software Updates

Parts

The following parts are needed:

• Raspberry 4 or compatible board, recommended 4GB or 8GB RAM, but even 1GB will work as memory consumption is about 700MB.
• Optional: Intel Neural Compute Stick (NCS) 2
• Waveshare Servo Driver Hat or compatible PCA9865 servo boards.
• USB camera with ~60 deg FOV, e.g. original DeepRacer camera or a Raspberry Pi camera.
• 4 RGB light diodes
• Misc cables

¹ The Waveshare hat comes with a step-down converter from 7.4V to 5V, and will power the Pi via the GPIO header. Other similar cards will require separate power.

Optional

• Fan/heat-sink kit for Raspberry PI 4

Software Install

Installation of software is reasonably straight forward, as pre-built packages are provided:

• Flash an SD card with Ubuntu 22.04 Server for ARM64 using the Raspberry Pi Imager.
• Boot the SD card, and let it upgrade (this takes some time...)
• Run git clone https://github.com/larsll/deepracer-pi
• Run sudo ./install-prerequisites.sh
• Reboot
• Run sudo ./install-deepracer.sh

Once installed you can start the stack with sudo /opt/aws/deepracer/start_ros.sh. To ensure a smooth start a camera needs to be plugged in. The launch log will now display in the console.

To automatically start at boot do sudo systemctl enable deepracer-core and to start the service in the background sudo systemctl start deepracer-core.

Changes

Some changes have been made to the code to enable access to GPIO as sysfs layout is different on the Raspberry Pi than on the custom Intel board.

Improvements

• Stripped down OS
• Runs the custom deepracer stack also seen in deepracer-scripts.

Details

PWM Outputs

Channel	Purpose	Notes
0	Speed controller	Remove red cable for stock DeepRacer ESC
1	Steering servo
2	RGB LED	Tail light - Red
3	RGB LED	Tail light - Green
4	RGB LED	Tail light - Blue
5
6
7	RGB LED
8	RGB LED
9	RGB LED
10	RGB LED
11	RGB LED
12	RGB LED
13	RGB LED
14	RGB LED
15	RGB LED

NOTE: Remove the red cable on the stock DeepRacer from the speed controller into PWM channel 0, otherwise you are putting 6V into the servo hat.

LiPo can power both the board and car, 3 pin (balance lead) gets wired to VIN (black and red cables only) to power the board and RPi. The 2 pin power cable goes to the car as normal.

GPIO layout:

• gpio1 - enables PWM (does not do anything, PWM is always on for the Waveshare board)
• gpio495-gpio504 - maps to PWM7 to PWM15 on the Hat, to control three RGB leds (those originally on the side of the board)

USB host mode: To connect the USB-C port on the Raspberry Pi 4 to a USB on your computer, then you need to perform two changes:

• Add dtoverlay=dwc2,dr_mode=host to your /boot/firmware/config.txt
• Add modules-load=dwc2,g_ether to your /boot/firmware/cmdline.txt

With this you will see a new network connect appear in Windows (and Mac?). The Raspberry Pi will have IP 10.0.0.1; you can connect to the console via https://deepracer.aws. This will require your computer to be disconnected from all other networks (given a DNS priority issue). SSH directly to the IP will work.

What does not (yet) work

• Battery gauge is not connected - red warning message persists
• Device Info Node is looking in non-existent places - no real impact