Inputs and Outputs of ONNX Model Deployment

[cooltech101]

Hi, I want to deploy COMPASS on a physical robot. I have successfully converted the model to ONNX format, but I don't know how to setup the inputs and outputs and was not able to find any documentation on this.

I understand that there are 6 inputs: image, route, speed, action_input, history_input, sample_input. And 3 outputs: action_output, history_output, sample_output. I believe all inputs and outputs are datatype numpy float32.

What kind of information do I feed into the 6 inputs and what is the format?

For action_input, history_input and sample_input, am I correct to assume that they are obtained directly from action_output, history_output and sample_output?

What about the image, route and speed inputs?

Thank you.

[nv-liuw]

Hi @cooltech101 ,

Thanks for your interest!

We don’t currently have inference examples using ONNX, but you can refer to the TensorRT examples in X-Mobility. The input and output setup is very similar, so they should give you a good starting point for preparing your own ONNX inference pipeline. https://github.com/NVlabs/X-MOBILITY/blob/main/ros2_deployment/x_mobility_navigator/x_mobility_navigator/x_mobility_navigator.py

Also, this PR in X-Mobility repo should give more info on the COMPASS inference integration in ROS: https://github.com/NVlabs/X-MOBILITY/pull/14/files#diff-5331a5f52eb30b6361db3878b7c028a4d5af5cf5f4fb7f4abd76692ac221cc75R203

[cooltech101]

Hi @nv-liuw ,

Thank you for your guidance! I was able to successfully implement COMPASS inference in Gazebo and will try it on a physical robot next.