Budget Bots Score Big

Autonomous systems like self-driving vehicles, humanoid robots, and drones are becoming much more intelligent, capable, and useful as related technologies — especially in the area of artificial intelligence — continue to advance. But we still have a long way to go. The control systems that power these devices tend to be brittle, frequently failing to perform as expected when they face challenging conditions. Moreover, the powerful computers required to run these algorithms are expensive and complicated to work with, which keeps them out of reach for many developers.

If we are going to solve these big problems and usher in a new era of intelligent machines, these hurdles must be overcome so that we can have all hands on deck. With more people working toward solutions, that day will arrive sooner. A pair of researchers at The University of Texas at San Antonio recently completed a survey of available technologies to determine the best way to run powerful computer vision algorithms on low-power, and relatively inexpensive, edge computing hardware. Their findings have the potential to make these technologies available to a wider range of developers.

In pursuit of this goal, the researchers worked to develop a low-cost, low-power embedded system equipped with a monocular or stereo camera that leverages machine learning and computer vision to detect and interact with objects. Ultimately, they hope that the system they design will help them with the 2024 International RoboCup competition by being able to locate, and interact with, a soccer ball.

The team utilized convolutional neural networks (CNNs) for object detection, which helped them to recognize and track soccer balls. The CNN architecture involved preprocessing the images, extracting key features, classifying objects, and predicting bounding box coordinates to locate the soccer ball in real-time. This information would enable a robot to act on the visual data effectively.

To support the system, the team experimented with two hardware options — the Arduino Nano 33 BLE Sense ML Kit and the Google Coral Edge TPU. Due to performance challenges with the Arduino kit, the Coral Edge TPU was selected for its faster inference time (30 ms) compared to a CPU (Intel Core i9-13900H 2.60 GHz, 240 ms) and a GPU (NVIDIA GeForce RTX 4070, 40 ms). This made the TPU an ideal choice for real-time object detection in a low-power, low-cost system.

The team further optimized the system by using cost-effective cameras. They tested both a stereo camera (Intel Realsense D35I) and a monocular camera, discovering that the latter provided comparable performance, for this particular task at least, which helped reduce overall costs without sacrificing detection accuracy.

Having landed on a successful combination of hardware and software, the researchers now intend to use it to power a humanoid robot that they can enter into the next RoboCup competition. Keep your eyes on this one to see how their inexpensive solution fares against more powerful hardware. Perhaps we will find that intelligent, autonomous robotic systems are more accessible than ever before.