MagicSim: A Unified Infrastructure for Executable Embodied Interaction

MagicSim is a new infrastructure for robot learning that integrates diverse simulation components into a unified framework, enabling executable embodied interaction through a deterministic batched runtime and a shared Markov decision process (MDP). It utilizes YAML-first specifications to create varied executable environments and supports a comprehensive command-to-robot execution pipeline, which includes benchmark evaluations, automatic trajectory collection, and interaction with visual language models. This unified approach enhances the reproducibility and efficiency of training and evaluating embodied agents, making it a significant tool for practitioners developing advanced robot learning systems.