Many people compare Raspberry Pi 4B with Jetson Nano. We have used two robotic arm products on different chips. One is mycobot280-pi based on Raspberry Pi 4B, and the other is mycobot280-JN based on Jetson Nano. Today we will use it to run an AI program to compare their performance, and want to see what the difference will be.
The two built-in systems are both Ubuntu18.04, and the rest are the same except for different motherboards. ROS and python are used to compile this environment. The main function of the AI Kit is the process of accurately grabbing wooden blocks of different colors by the robotic arm under the recognition of the camera, and placing the colored wooden blocks in the buckets of the corresponding color.
First we set up the AI kit and connected to the robotic arm, and then ran the program.
We ran the same program twice with different robotic arms.
First, we ran the vision.launch file to create a 1:1 environment.
Then run the color capture file detect_obj_color.py, and then the robotic arm will recognize and capture the wooden blocks under the camera.
We found that when recognizing images, the procession time of the Raspberry Pi robotic arm is usually 3-4s, and then the arm grabs the colored wooden blocks, while the image recognition of the myCobot Jetson Nano is usually completed within 1-2s, then the arm would grab the colored blocks.
Here is the comparison of Raspberry Pi 4B and NVIDA Jetson Nano:
The VideoCore VI used by the Raspberry Pi 4B is a classic dual Shader architecture, with a performance of only a few dozen of GFlops, while the GPU performance of the Jetson Nano can reach 472 GFlops. The gap between their own GPUs is relatively large.
In the less standardized evaluation of two different chip robotic arms, we can easily find that the GPU of Jetson Nano is much more powerful than that of Raspberry Pi.Its performance in all aspects is relatively stronger.
I have two more questions, one is whether there is any other way to compare the characteristics of the two motherboards, and another is whether it is possible to improve the speed of image recognition by modifying the code without changing the hardware on top of the original one?
Whole project file:
Here we run the /launch/vision.launch and /scripts/detect_obj_color.py files.
AI Kit Videos：