Is it mandatory to use Jetson AGX Orin 64GB for robot deployment? What is the minimum configuration? Are there any cheaper options？

[CRONOIST]

Hi, your work is great!
Is it mandatory to use Jetson AGX Orin 64GB for robot deployment?
What is the minimum configuration?
Are there any cheaper options？
Thank you！

[sunsleaf]

Hi, Jetson AGX Orin 64GB isn't strictly needed for on-board deployment. The Orin NX from Unitree Go2's expansion dock could also be enough, although we haven't tested it.

Orin NX should run the policy networks fast enough. For the perception module, if the inference is too slow, you can try to reduce the number of layers and heads for the Transformer, or reduce number of layers or channels for the U-Net, which should give significant reduction on the network size and make it run more smoothly on-board.

[CRONOIST]

Hi, Jetson AGX Orin 64GB isn't strictly needed for on-board deployment. The Orin NX from Unitree Go2's expansion dock could also be enough, although we haven't tested it.

Orin NX should run the policy networks fast enough. For the perception module, if the inference is too slow, you can try to reduce the number of layers and heads for the Transformer, or reduce number of layers or channels for the U-Net, which should give significant reduction on the network size and make it run more smoothly on-board.

Okay, thank you very much for your answer！I will try it on the expansion dock of Unitree Go2. I wish you good health and smooth progress in your scientific research!

[CRONOIST]

Hi, Jetson AGX Orin 64GB isn't strictly needed for on-board deployment. The Orin NX from Unitree Go2's expansion dock could also be enough, although we haven't tested it.

Orin NX should run the policy networks fast enough. For the perception module, if the inference is too slow, you can try to reduce the number of layers and heads for the Transformer, or reduce number of layers or channels for the U-Net, which should give significant reduction on the network size and make it run more smoothly on-board.

Hi, I have logged into the expansion dock of unitree go2 and found that the Ubuntu system is 20.04, and ROS2 can only install the foxy version, not the humble version. Can I complete the deployment by following the open-source deployment process? Looking forward to your reply again! Thank you！

[sunsleaf]

I think the code doesn't depend on humble-specific features. Haven't tested it though. You can have a try.

[sunsleaf]

Hi, you need to use a USB-C expansion dock with an ethernet port to have one more network interface in addition. We connected the LiDAR to this exoansion dock and the Robot to the builtin ethernet cable on Jetson

[CRONOIST]

Hi, you need to use a USB-C expansion dock with an ethernet port to have one more network interface in addition. We connected the LiDAR to this exoansion dock and the Robot to the builtin ethernet cable on Jetson

Hi, thank you very much for your patient and professional response! I have successfully deployed REASAN on a laptop equipped with a GTX 4060. This is an impressive piece of work, and its obstacle avoidance function is excellent! However, when testing the navigation function, I found that go2 cannot quickly and stably reach the designated target point. It often circles in place and cannot stand steadily after reaching the target point. Is there any way to improve this? Looking forward to your reply again! Wish you good health and smooth work!

[sunsleaf]

This is largely due to the bias introduced during training the navigation policy. Here are some ideas you could try.

For staying at the goal:

• Enlarge the noise for the ray perception when training the navigation policy to increase stability in real-world (noisy perception).
• Increase the reward weight for staying at the goal.
• Remove the moving obstacles when training the navigation policy.

For reducing circling:

• Increase the hidden dimension of the LSTM to provide more memory ability.
• If your testing scenarios are simple (not containing long detours), you can train only on the first type of scenes ("general"). Or you can design the training environments to match your testign environment better.
• Also remove the moving obstacles when training the navigation policy.

Hope this helps. Would be very glad to hear if you get any improvements!