Do I need two SO 100 arms to get started?

[rabhishek100]

I have printed and assembled one arms, the follower version. Do I need two arms to record datasets and do testing?

[JohnsonLv]

you need a leader to tell follower what to do

[rabhishek100]

can i use the same arm as the leader to record datasets and as a follower while running policies?

[fracapuano]

Hey @rabhishek100 👋

Short answer

Nope

Long answer

• Assuming you are training policies using imitation learning directly from frames (that is, using ACT and thus following the tutorial quite verbatim),
• Assuming you intend to directly manouver the sole (follower) arm while recording

then you would be in the camera frames when imitation-learning, which would make learning impractical. You could be using one arm only in the case where you were demonstrating manouvering the robotic arm, but you would have to learn from sensors (in the case of the arm, joint angles + your own sensors), which I think is a bit beyond the current scope of the arm. For that, I think you would be much better off 3D-printing the other arm and using teleoperation to record your dataset instead of manouvering.

In case you are not learning from pixels, but rather from sensors, hmu! I would love to learn more about what you are doing and perhaps help you if needed 😊

[rabhishek100]

Hi @fracapuano, thanks for the response.

I understand what you mean. We cannot have our hands in the frames while learning. When I started off with this, I had a few free servos laying around and I was able to assemble an arm. I was thinking of a way to do imitation learning with just the one arm. I played around with the arm and was able to teleoperate it pretty well using just a keyboard, by doing inverse kinematics. But later, I got around to building a second arm and successfully trained act for a simple pick and place task. It would be great if we can use keyboard teleop for training the arm (we can read the actions directly from the follower arm itself). This way, people can easily get started with just a single arm, reducing the time and cost by half. Let me know if I am missing something here.

This is the keyboard teleop I was able to achieve - https://www.youtube.com/shorts/emt-0YKsKAA

[fracapuano]

@rabhishek100 great! Would you mind putting here the inverse kinematics code you have produced? I would gladly give it a look 👍

[rabhishek100]

Sure! I have attached the files
keyboard_teleop.zip