MAGMA Documentation

MAGMA is a framework for embodied AI agents working on interactive, long-horizon, multi-robot tasks. Use provided scenarios or build your own tasks and environments to generate training data with MAGMA-GEN.

These guides cover the stable v2 packages: MAGMA-GEN 2.0.0, Core 2.0.1, and Scenarios 2.1.1. Start with the installation guide for a compatible setup.

Choose your starting point

The choice is yours.

blue pillGenerate training data ↗MAGMA-GENExplore your agent’s decisions, test corrections, and export training examples.Start generating →red pillEvaluate your model ↗MAGMA-BENCHInstall the required packages, download the benchmark, and evaluate your agent on interactive robot tasks.Explore the benchmark →

Pick your next move.

You can also check the developpers guide to create your own tasks and scenarios or connect your own model or agent.

New to magma?

Start with Play a Task to install the scenarios and control a robot with the tool tester.

Understand and look things up

• Concepts and architecture explains scenarios, tasks, stages, tools, environments, and services.
• Reference collects scenario authoring and agent integration contracts.