OpenWAM: An Open, Modular Exploration Towards Systematic World-Action Model Pretraining

OpenWAM. An Open, Modular Exploration Towards Systematic World-Action Model Pretraining. By Yuran Wang and twenty-three others, from the National University of Singapore, Tsinghua University, and five other institutions. Published on the seventh of September. Its code repository is worked out: about four hundred and fifty source files, with a readme and a licence.

World-Action Models inherit knowledge of the world from video-generative priors and channel it, through embodied experience, into executable control signals. Existing systems are monolithic. The generative backbone, the visual representation, the architecture, the information flow, the inference procedure, and the training data are all tightly coupled, which hides which design choices matter and why. When everything changes together, a better result cannot be traced to any one decision. OpenWAM is an open research stack that turns world-action pretraining into a controlled experimental program, and its study distills three principles.

First, upstream knowledge transfers through a sufficiently capable generative backbone and a compact, information-rich latent space. Second, synergy between world and action learning requires dedicated action capacity, explicit information flow from world to action, and synchronised joint denoising. Third, embodied pretraining mainly improves out-of-domain generalisation, and one-stage co-training over egocentric and robot data integrates coverage of the world with grounding of actions. Read together, the principles say where knowledge comes from, how the world half and the action half must be wired to help each other, and what embodied data adds on top.

The stack has two parts. OpenWAM-Infra factorises the design space into composable modules, with unified training, inference, deployment and evaluation. On that substrate, OpenWAM-Study runs controlled experiments on three questions: what to inherit, how world and action learning interact, and how their synergy scales. Those experiments are the source of the principles. Because each module can be swapped while the others stay fixed, a difference in outcome can be attributed to the module that changed, which is what makes the study controlled.

Those principles are then composed into a model, OpenWAM alpha, pretrained on roughly six thousand four hundred hours of egocentric human and robot data. It is evaluated on eight simulation benchmarks and in real-robot experiments, spanning embodiments from single-arm and bimanual manipulation to dexterous hands. Across all of them it delivers consistently excellent performance, keeping its top-tier standing from simulation into the physical world.

The principles come from the design space the modules cover and the questions the study asked, and the third principle is specific about what embodied pretraining improves: principally generalisation out of domain. The model's standing is reported on the eight simulation benchmarks and the real-robot experiments described. The full stack is released, including infrastructure, evaluation protocols, pretrained models, and data recipes, so the experiments can be extended.

The contribution is as much a way of working as a result. Decoupling a world- action model into swappable parts makes it possible to ask which part matters, and the three principles are what that questioning produced. The released model is those principles put together.