oniro
Reinforcement learning trainers compiled to WebAssembly. Each demo trains in your browser, on your machine; nothing runs on a server.
CPU demo
REINFORCE, PPO, SAC, Dreamer and TD-MPC2 on the byte-exact CPU backend, rendered with WebGL2, which current browsers support. Its "Watch pretrained" links replay three tiny trained policies with training disabled.
WebGPU demo
GPU-resident training on WebGPU. Needs a browser with WebGPU enabled (current Chrome or Edge). On a software adapter it warns, because training then runs on the CPU.
A run can start straight from the URL, for example
cpu/?algo=ppo&env=cartpole&seed=1&iters=200&go=1, and a shipped
policy can be watched the same way:
cpu/?algo=ppo&env=acrobot&policy=/oniro/cpu/policies/acrobot.policy.json&watch=1&go=1.