How Nature Solved Robotics

This video is both a story of my adventure with AI robotics and the fascinating lessons I learned along the way. Try Mammouth AI now at http://mammouth.ai

I called this project GrowBot. Raspberry Pi Zero 2, two servos, a camera, an IMU, a battery, some 3D printed parts. I trained small neural networks in simulation so it could walk, stand and spin in a lifelike way. Then I handed control of the motors to a vision language model. It read its raw sensor data, wrote its own code, built profiles of the people it met, and dreamed between sessions to clean up what it had learned.

It worked better than I expected, until it hit a wall. Which kept pointing back to the same question. How do you act smoothly when thinking is slow? That sent me into how nature solves the problem (the cerebellum), and it turned out to be exactly what the robotics field is converging on right now.


0:00 The robot that didn't want to be alone
1:15 Meet GrowBot
2:21 Building the $80 body
3:51 Sponsor
5:01 Learning to move (System 1)
7:05 First magic moments
8:20 Handing the controls to AI
11:38 Personality and Disney mode
15:03 Memory, agents, and dreams
17:36 The wall: physical dumbness
19:23 Nature's trick: the cerebellum
23:23 DayDreamer and the converging architecture
26:20 Alpha test

#machine #Robotics #robot