coolspeed

coolspeed · AI engineer · MLOps engineer

Game characters that learn to move

This site has Gold Spider, a robot that learned to walk and runs live in your browser, and the dev log about how I made it.

Try the Gold Spider demo Read the dev log
Project

Gold Spider

Live demo

A six-legged robot that learned to walk through reinforcement learning. Throw a 171 kg steel ball or a 79 kg crate at it and watch it stagger, recover and keep walking over rough block terrain.

480Hz
physics steps per second
30Hz
neural network decisions per second
24
joints the network drives
0
hand-made keyframes
Open the demo How it was made

Runs on your own device, desktop or phone. About 12 MB on the first visit.

Writing

Dev log

All posts
About

coolspeed

AI engineer and MLOps engineer.

I started out as a game server programmer, working on distributed systems and networking. Now I train models and run them: reinforcement learning for game characters, and LLM inference, locally and in the cloud.

What I work on

  • Reinforcement learningTraining policies for physics-based characters. Gold Spider is the latest.
  • MLOpsRunning training and inference, locally and in the cloud, and making it fast.
  • Game serversWhere I started: distributed systems and networking for online games.