Waypoint Labs
Agents that learn by doing.
Waypoint Labs is a fictional research studio that explores reinforcement learning: software agents that improve through trial, error and reward. Watch one learn, read how it works, and see how we think about doing it responsibly.
Four ways in.
Start with the live demo, or read first and experiment later.
Playground
Train a small agent to cross a gridworld, right in your browser, and watch its policy improve.
Open Playground
Research
Rewards, exploration, policies and discounting, explained in plain language.
Open Research
Approach
Safe exploration, honest evaluation and human oversight, and how this site follows them.
Open Approach
About
The student project behind Waypoint Labs, how it is hosted, and common questions.
Open AboutThe idea
Reward the outcome, not the steps.
In reinforcement learning nobody tells the agent which move is correct. It tries actions, receives a reward or a penalty, and gradually learns which choices lead to good results over time.
The same loop powers game-playing systems, robot controllers and recommendation tuning. Here it is small enough to watch in a few seconds.