Oniro

Oniro is a reinforcement-learning trainer in Rust, built on the Burn deep-learning framework. It trains on the environments of gymnasium_rs and runs natively, from Python or in the browser.

Live demo

The Oniro playground after a PPO run on CartPole, with its rising learning curve
Open the live demo. Pick an algorithm and an environment, and watch the learning curve rise.

Features

  • Algorithms: REINFORCE, PPO, SAC, TD-MPC2 and a Dreamer-style world model, plus DQN on the GPU.
  • Reproducible: CPU training is deterministic for a given seed.
  • Interfaces: a command line with TOML configuration and checkpoints, Python bindings, and browser builds for the CPU and for the GPU through WebGPU.

Status

Oniro is in pre-release development; its source is not public yet.