noeira

• 350+environments and tasks
• 35+learning algorithms and planners
• 0Python between training and the motors
• 1,100+test files

From the simulator to the robot, in one language

A robot-learning project today is assembled from pieces that were never meant to meet: a simulator for physics, a deep-learning framework for networks, a robot-learning library for datasets and policies, a vendor SDK for the motors, and whatever runtime fits on the robot’s computer. Every seam costs something — copies across language boundaries, versions that drift, a policy that behaves differently on the robot than it did in training because the code that ran it was not the code that trained it.

noeira builds the whole loop in Mojo instead: the physics engine, the task suites, the neural networks, the learning algorithms, the camera pipeline, the servo bus and the deployment runtime. The policy that trains on a GPU in the cloud is the same code that drives the robot from its onboard computer.

simulationlearningthe real worlddata

What is inside

Simulation

A MuJoCo-inspired 3D engine checked against MuJoCo itself — meshes, tendons, contacts, sensors — with batched GPU stepping, ray-traced cameras and a library of environments and robot models. See 3D physics and environments.

Tasks and benchmarks

A declarative task layer — families, goals, placements — that describes a task once and runs many of them in one batched GPU scene. Benchmarks such as LIBERO are ported onto it. See the task layer.

Learning

Reinforcement learning from tabular methods to SAC and PPO, world models (TD-MPC2, DreamerV3, Dreamer 4), MCTS planners, and zero-shot Forward-Backward. See algorithms.

Imitation and VLAs

Learning from demonstrations: behaviour cloning, action-chunking transformers, vision-language-action models, and reinforcement learning with human interventions — from data recorded in simulation or on a real robot. See imitation and VLAs.

Robots

Hardware in the loop: servo buses, calibration, teleoperation, cameras, recording, and policies deployed on edge GPUs with the safety rules built in. The first supported robot is the SO-101 arm on a Jetson Orin. See robots.

Infrastructure

noeira.nn — gradients composed at compile time, 70+ primitives, fused and flash attention, AMP and CUDA graphs — plus native HDF5, Parquet and video I/O, datasets in the LeRobot v3 format with no dependency on LeRobot, and noeira cloud for runs and datasets.

Start here

Install

Set up pixi and run a first example. No GPU required to get going.

Train in simulation

Train DQN on CartPole on the GPU, or start smaller with Q-Learning on GridWorld.

Bring up a robot

Set up a robot — the SO-101 is the first supported — record demonstrations, and deploy a policy, safety rules first.

New to the field?

RL in five minutes — the agent and environment loop, and the vocabulary the rest of the docs assume.

Understand the design

The stack and compile-time composition, the idiom the rest of the code is built on.

Project status

noeira is developed in the open and was selected for a Modular Community Grant in 2026. Maturity differs by area, and every page carries its own marker:

• Stable: the 3D physics engine and its MuJoCo validation, the classic and MuJoCo-style environments, DQN / PPO / SAC / TD3, noeira.nn.
• Beta: the task layer and its benchmarks, imitation learning on a real robot, dataset recording, deployment on edge hardware.
• Experimental: vision-language-action models on a robot, learning from interventions, the world-model agents, DeepMind Control training. Results are reported with their caveats on each page.