OpenWAM, a new research paper shared on Hugging Face, proposes an open and modular approach to pretraining world-action models (WAMs). The authors argue that existing systems are monolithic: the generative backbone, visual representation, architecture, information flow, inference procedure, and training data are tightly coupled, which makes it difficult to determine which design choices actually matter.
To address this, the work introduces three components. OpenWAM-Infra factorizes the WAM design space into composable modules with unified training, inference, deployment, and evaluation across eight simulation benchmarks. OpenWAM-Study uses controlled experiments to examine what upstream knowledge to inherit, how world and action learning interact, and how their synergy scales. From these experiments, the authors distill three principles: upstream knowledge transfers through a sufficiently capable generative backbone and a compact, information-rich latent space; world-action synergy requires dedicated action capacity, explicit world-to-action information flow, and synchronized joint denoising; and embodied pretraining mainly improves out-of-domain generalization, with one-stage co-training over egocentric and robot data integrating world coverage and action grounding.
Building on these principles, the team presents OpenWAM-α, an open WAM pretrained on roughly 6,400 hours (518.5 million frames) of egocentric human and robot data. The model is evaluated across eight simulation benchmarks spanning five embodiments and on real-robot experiments covering single-arm and bimanual manipulation as well as dexterous hands.
According to the paper, OpenWAM-α achieves top-tier results across all eight simulation benchmarks, including state-of-the-art performance on the mobile bimanual benchmark EBench. In real-world evaluations, it ranks first on the bimanual RoboDojo-Real leaderboard, reportedly doubling the success rate of pi0.5, and consistently outperforms representative WAM and VLA baselines on single-arm and dexterous-hand robots.
The authors release the full stack—infrastructure code, evaluation protocols, pretrained and post-trained weights, and data recipes—with the aim of providing a strong, reproducible baseline for future research. Project resources are available via the project page, GitHub, and Hugging Face.