Open-WAM: Open-source World Action Models Libraries

OpenWAM: Two Open Projects Rethinking How Robots Learn to Act What Is a World–Action Model? Most robot AI systems maintain two separate modules: one that imagines what might happen next (a world model) and another that decides what to do (a policy). These are typically trained independently and then combined. World–Action Models (WAMs) challenge this separation. The core idea: video and action are two streams of one sequence, not two disconnected systems. Actions create visual futures; visual futures give meaning to actions. A WAM learns both together, letting each inform the other. --- Project 1 — The Stanford SVL Infrastructure Framework Source: openwam.stanford.edu · github.com/OpenWAM/OpenWAM Developed by the OpenWAM Team at the Stanford Vision and Learning Lab (SVL), this project is an extensible, modular framework for video-action world models in robot learning. What It Does The framework's central design principle is separation of concerns. Rather than bundling everything into one monolithic system, it factorizes the WAM design space into composable, interchangeable parts: Model topology — how the transformer is structured Video/action conditioning — how inputs are fed to the model Sequence semantics — how the token sequence is interpreted Visual execution — how observations are processed Action decoding — how action outputs are generated This separation means that controlled experiments can share the same trainer and visual stack, making previously incomparable…