AI Takes Pole Position: Enthusiasts Adopt Jetson Nano for DIY Robocars Race

The small but mighty developer kit is accelerating hobbyists’ race cars and setting track records.
by Scott Martin


Stock cars zigzag to the finish line. Spectators cheer. Pit crews high-five. It’s not NASCAR but a hardware co-working space that’s ground zero for the DIY Robocars race car movement in Silicon Valley.

In place of mechanics with wrenches are tech industry engineers and hobbyists sporting laptops to fine-tune their neural networks.

On this scene, NVIDIA Jetson Nano Developer Kit offers a performance leap into the autonomous arena. That’s because these enthusiasts — many racing in the 1/16-size or smaller “stock” car category — are discovering it offers professional AI on an amateur’s budget.

The robocar races draw in students, hobbyists and off-duty employees from Amazon, Google, Microsoft, NVIDIA and elsewhere in tech for weekend fun. “NVIDIA has become more relevant for our community,” said event host Chris Anderson, CEO of drone startup 3DR. “This is a proxy war for full-size autonomous vehicles.”

On a recent Sunday in September, a couple hundred enthusiasts gathered at the Circuit Launch space in Oakland, Calif., for a robo race. Just months earlier, many in this maker community of do-it-yourself racers were using Raspberry Pi, but now they’ve mostly switched over to Jetson Nano to stay competitive.

NVIDIA’s team took first place in the “stock” car race. In the “unlimited” race, which allows bigger-budget cars, Jetson was used in the cars that placed first and second.

More than 10,000 people across 40 countries are involved in the roborace community worldwide, according to Andersen.

Faster Track Times

Not long ago it wasn’t this way. At the start of the year, record track times were nearly double what some have been posting lately. Cutting their time around the indoor track — hitting roughly six-second times versus 12 seconds — has been a payoff of deep learning from Jetson Nano.

“Jetson Nano is definitely starting to be adopted as a maker platform. It’s an example of this paradigm shift to AI and how this maker community is taking the leap,” said John Welsh, developer technology engineer for Autonomous Machines at NVIDIA, who was in the race.

Why Jetson Nano AI

The compact and powerful Jetson Nano, which runs Ubuntu Linux, operates the same AI frameworks — including PyTorch, Keras, TensorFlow and OpenCV library — used on supercomputers.

PyTorch open source machine learning libraries for computer vision and natural language processing allow bots from all walks to see and talk.

Jetson Nano also runs the NVIDIA CUDA-X collection of libraries, tools and technologies that can boost performance of AI applications. This means it can use all the same TensorFlow software libraries and can enable deep learning to optimize models and speed inference with TensorRT.

Plus, Jetson Nano delivers 472 GFLOPS of compute performance to handle demanding AI workloads while running on as little as 5 watts, ideal for battery-driven bots of all types.

“It’s amazing to be able to bring this autonomous technology to everybody,” said Gabriele Gorla, Jetson architect at NVIDIA, who was also at the race.

Build Your Own DIY Racer

If you want to get started, the JetRacer open-source project provides a great starting point to tackle races with AI. Just pick up a JetRacer car, which includes Jetson Nano, and begin assembly. A bill-of-materials for ready-to-build JetRacers can be found here and here. Assembly instructions are here and here.

Another way is to build your own customized car from scratch —  or consider a DIY Donkeycar —  and fit it to Jetson Nano.

To speed software development, flash the available JetCard files on a microSD card. This system configuration comes preloaded with a Jupyter Lab server that fires up at boot for web-based programming. It’s helpful to include an LED display in your build to show the Jetson’s IP address and other useful information.

The system setup also provides access to the previously mentioned and popular deep learning frameworks of PyTorch and TensorFlow. After configuring your setup with JetCard, you can start developing AI projects straight from a web browser in Python.

Robo Racing and More

The open source Ubuntu Linux enables powerful programming for deep learning. Using the Jupyter Notebooks, hobbyists and professionals alike can easily train models. It’s as easy as changing a block of code in a web browser, and then continuing to use the model.

Robo-racing is just the start. Potential applications for Jetson Nano include everything from medical imaging to home robots and industrial IoT. With Jetson Nano, the scope of AI development to undertake is only limited by the imagination of users. A list of projects under development within the Jetson community can be found here.

“We believe there’s much more potential with using deep learning,” said Chitoku Yato, technical marketing manager for Jetson at NVIDIA, who was on the race team with Welsh. “That’s why we’re demonstrating the platform for other users.”