Seven Universities Open-Source OpenWAM, a World-Action Model for Robots

On September 24, research teams from seven universities, including the National University of Singapore, Tsinghua University, and Peking University, jointly open-sourced OpenWAM, an open-source research stack and foundation model for World-Action Models (WAM).

OpenWAM consists of three components: OpenWAM-Infra, a modular infrastructure; OpenWAM-Study, a study of design principles; and OpenWAM-α, the foundation model. The team also released the paper, code, models, and data to provide a reproducible common starting point for WAM research.

The core idea of OpenWAM is to combine video generation models with action learning, enabling models to understand how environments evolve over time while learning how to change the future through actions. The research decomposes the problem into three levels: what knowledge to inherit, how world modeling and action learning can cooperate, and how this cooperation scales across data domains and embodiments.

OpenWAM-Infra defines composable models, training runtime, deployment runtime, and evaluation protocols, supporting single-system, dual-system, and triple-system architectures. OpenWAM-Study derives key design principles through controlled experiments: strong generative world priors, compact and information-rich visual representations, explicit world-to-action information flow, and cross-domain embodied pretraining are critical for building high-performance WAMs.

OpenWAM-α is built on these principles, using Wan2.2-TI2V-5B as the video stream backbone and a 1B-parameter ActionDiT for the action stream, with an 80-dimensional unified action space to support multiple robot morphologies. Pretraining data includes human egocentric videos, real robot data, and synthetic robot data, with the actual training set comprising approximately 518.5 million frames (about 6,369 hours).

On eight simulation benchmarks, OpenWAM-α achieves state-of-the-art results on EBench and demonstrates cross-embodiment generalization in real robot experiments. The team noted that future work requires larger-scale, more diverse embodied data with precise action labels and more robust visual representations.