Alloy physics engine released for robot learning

Researchers introduce Alloy, an open-source physics engine designed for robot learning that enables rigid robots, cloth, cables, liquids, granular materials, and soft bodies to share a single simulated scene. The engine assigns objects to specialized material solvers, couples them during control steps, and operates in parallel across multiple GPU or CPU worlds while providing health grading to distinguish physics failures from policy errors. Built on Newton, MuJoCo Warp, and Warp, Alloy is integrated with the Model Context Protocol to allow coding agents to install and train policies for tasks like walking, cable handling, and pouring. Open article →