1
$\begingroup$

I'm in the process of learning Tensor Flow, and I really want to use it on a new robot I'm going to build. I want the robot to be able to do image recognition, and move towards an image of interest, or follow a path, using neural networks. Is this even possible with tensor flow?

Is the Arduino capable of running neural network frameworks, such as Tensor Flow?

Thanks.

$\endgroup$
2
  • $\begingroup$ All of the things you want to do is possible with opencv. It would take a long time and very large datasets to achieve what you want with tensorflow $\endgroup$ – Ozymandias Apr 19 '18 at 18:56
  • $\begingroup$ Oh ok thanks ill look into that. Do you have any recommendations for parts? In particular the board, and vision related $\endgroup$ – Remus Ventanus Apr 19 '18 at 21:34
2
$\begingroup$

The Arduino is a microcontroller and not a "full" computer. (See my other answer here.) As such, you will have a difficult time running TensorFlow on the Arduino.

Deploying TensorFlow, or another Neural Network Framework, on your robot has at least two parts:

  1. Training: teaching your network, best performed on a computer or server with a powerful GPU
  2. Deployment: loading your trained model onto your platform for realtime operation.

You would have an easier time training on a desktop computer and deploying to a robotic platform that uses a Raspberry Pi for high-level tasks and an Arduino for low-level tasks. On the Pi, you could use TensorFlow or OpenCV in either C/C++ or Python.

Depending on your objectives, you may be able to get away with using specialized camera hardware like the Pixy, which would work directly with the Arduino.

$\endgroup$
3
  • $\begingroup$ Hi, and thanks for the response. My real objective is just to get my feet wet with intelligent robotics. Ive done AI, but robotics is the area I really want to get into once done university. The project I have in mind is a robot that can use either the genetic alg, or NN to perform pathfinding. I may use A* first as I believe it will be easier to implement. Im just sort of overwhelmed on what the best robot for me is $\endgroup$ – Remus Ventanus Apr 20 '18 at 1:30
  • 1
    $\begingroup$ If you’ve got the experience, jump right in with the Jetson TX2. You can either use TensorFlow or TensorFlowRT on that dev board. Personally, I find the Pi painfully slow for anything too intense. If you’re looking for a mechanical platform, check out the race car from Jetson Hacks. $\endgroup$ – Mepix Apr 20 '18 at 1:35
  • $\begingroup$ Im not too experienced with robotics, but have moderate with AI. Ill definitely look at the TX2. In your opinion, is A* on an arduino a reasonable starting point? Ive done A* for maze problems, so I feel it might be pretty simple to take the current code I have and apply it to a robot for pathfinding $\endgroup$ – Remus Ventanus Apr 20 '18 at 2:01
0
$\begingroup$

Check out an app on Google play (Arduino Object Detection and tracking by Gemcode Studios) It uses both opencv and tensor flow and sends image details over Bluetooth to an Arduino for further robot control and processing.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.