AI Battery Assembler

Inspiration

Garrett and I both interned at SpaceX in 2025 building industrial robots for Starship.
For this project, we wanted to apply our learnings from industrial robots to create practical machine learning applications.

What it does

Our project does three things:

1. LeRobot finds the orange cell in its view
   
2. LeRobot picks it up
   
3. LeRobot places it

How we built it

1. We built the physical cell
   
2. We got everything "talking" on our network — cameras, both LeRobot devices, MiniPC, and DevCloud via SSH connection
   
3. We ran training runs varying the position and orientation of the battery cell, as well as the number and location of batteries, to train the robot on all possible scenarios

Challenges we ran into

1. Robot calibration issues — getting tele-operation mode to run
   
2. KR260 firmware issues (had to swap units)
   
3. Acquiring consistent, highly repeatable training data for our models

Accomplishments that we're proud of

1. First hackathon for both of us
   
2. First time training an AGT model and deploying it
   
3. First time setting up custom Ubuntu networks on multiple different IPs and subnets at the same time
   
4. First time using LeRobot and training/deploying custom models on Hugging Face
   
5. Training run did not crash overnight

What we learned

1. Only the paranoid survive
   
2. Machine learning notices all patterns — including the stuff in the frame you forgot was there
   
3. Always ask all the questions you thought of asking

What's next for AI Battery Assembler

1. FPGA accelerated precision motor control
   
2. More robust training data — 50 training sessions isn't enough for mass production
   
3. Scaling AMD compute to accelerate our model training cycle times

Built With

• amd
• devcloud
• lerobot
• python