Stack-chan has plenty of personality, but I wanted to give it a way to move around. So I combined an M5Stack Stack-chan with a LEGO motor hub and wheels and explored whether it could learn to drive from demonstrations.
The idea was simple: drive the robot myself, record examples, and train a policy to reproduce the behavior. LEGO made it easy to experiment with the moving base, while Stack-chan provided the camera—and the face!
Here’s the video, including the build, data collection, driving trials, and wrong turns:https://www.youtube.com/watch?v=_1pQTt8gqZM
I started by assembling the LEGO base and mounting Stack-chan on it. Then I collected driving demonstrations, pairing camera observations with driving actions. I used these recordings to train an Action Chunking Transformer, or ACT—a policy that predicts a sequence of actions from its observations.
During supervised trials, the robot learned to follow a black line. With a separate set of demonstrations, I also tested driving between red-and-white guardrails without a line. These were small, controlled experiments, and the mistakes were an important part of figuring out what needed improving.
For the software and training setup, I used AI coding agents with robotics skills from Robium, an open-source repo I’ve been developing. Its LeRobot and Hugging Face skills helped guide the dataset and training workflow. This project also gave me concrete problems and lessons to feed back into those skills.
My favorite part was seeing a little desktop robot become something I could teach through driving examples. There’s still plenty to explore, including different tracks, lighting conditions, and more varied demonstrations.
The build write-up links to the reference app and code if you want to explore the setup or adapt it for your own robot:
Tutorial and project code: https://robium.ai/blog/lego-stackchan-drive/Robium robotics skills: https://github.com/robium-ai/robium
Looking forward to hearing what you think and seeing what others build!
Robium








Comments