Intern Robotics has unveiled InternW0-Δ, a World Action Model (WAM) designed to jointly model visual dynamics and robot action generation for generalist manipulation. The work addresses a central difficulty in the field: how to bring together priors from large-scale pretrained models — visual dynamics, scene semantics, geometry, and motion — inside one coherent framework for producing robot actions.
The architecture relies on a Mixture-of-Transformers (MoT) design in which a pretrained video expert and an action expert interact under semantic guidance from a frozen vision-language model. A pretrained 4D foundation model injects geometric and motion priors through training-only distillation, allowing the system to benefit from those signals without carrying them into inference.
The team also introduces Causal Imprint, a mechanism that learns future-relevant scene changes from training-only future supervision and feeds predictive representations directly to the action expert — avoiding the need for future-video rollout at inference time.
To support large-scale joint training, the researchers assembled a heterogeneous corpus combining robot demonstrations, UMI data, egocentric human demonstrations, and Ego2Robot data, curated and aligned under a common state-action representation. According to the paper, the resulting corpus contains over 20,000 hours of processed training data, which the authors describe as the largest open-source corpus of its kind to their knowledge.
InternW0-Δ was pretrained on this corpus and is reported to outperform prior methods across simulation benchmarks and real-robot platforms. The team states it will open source the training code, model weights, infrastructure, data-processing pipeline, and processed data where licenses permit.