Train ACT with LeRobot (Action Chunking Transformer tutorial)

preview_player
Показать описание
In this video, I show you how to train a robotics AI model for the SO-100.
The SO-100ARM is a fully open-source robotic arm project launched by TheRobotStudio, with accessible parts. It's a very popular arm in the AI robotics community, with lots of datasets and a big community.
This step by step guide explains how you setup your robots, your cameras, and your recording to create a high quality dataset. Then, it show you how to train and run ACT (Action Chunking Transformer) in order to get an AI robotics model. Finally, it shows how to control your robot with AI. Be careful and make sure your setup is safe.

00:07 Intro
00:52 Learning about ACT
02:23 Set up your robots
02:54 Set up the recording scence
04:08 Install and set up phosphobot
05:29 Sync your data with Hugging Face
06:36 Calibrate your robot and adjust recording settings
09:01 Control your robot
12:02 Explore your datasets
12:49 Train your ACT model
18:19 Control your robot with AI (inference)
23:09 Final advice and demo time
Рекомендации по теме
Комментарии
Автор

the model can be trained in the cloud, but what hardware do we need for inference?

tektronix
Автор

My god this is exactly what I wanted to build myself! Thanks a lot! I will check this today! 😊

swannschilling
Автор

Thank you so much. It's such a great tutorial. 2 questions,
1- The generated dataset uploaded to Hugging Face only has the image and the label 0 or 1 now. Is that the new data structure?
2- My cameras are set up so that the arm camera is the primary and the scene camera is the secondary. Does it matter? And how do I configure it to switch it if needed?

CodeWithAhmedM
Автор

what camers are you using for training?

TheRealJoshuaYang
Автор

Did you have to scale down resolution? I'm finding 1920 x 1080 x 3 is WAY too large for training

TheRealJoshuaYang
Автор

Can I use Macbook pro M1 with 16GB to reproduce your work?

Leo-Guo
Автор

super ! could you please give me the M3 config ? which "GPU/MPS" is it to have an idea ? thanks for your video really usefull

jean-marieprigent
visit shbcf.ru