InternW0-Δ: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data

InternW0-Δ. A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data. By Xingyu Miao and forty-seven others, from the Shanghai AI Laboratory. Published on the twenty-fifth of September. No code repository is linked to it.

World Action Models, or WAMs, jointly model visual dynamics and action generation, with generalist robot manipulation as the goal. A central challenge is integrating priors from large-scale pretrained models into one framework that generates robot actions. Those priors are of several kinds: visual dynamics, scene semantics, geometry, and motion. Each comes from a different kind of pretrained model, and the difficulty is making them work in a single unified system.

InternW0 Delta is a unified WAM, pretrained on a heterogeneous corpus. It combines pretrained visual dynamics, scene-level semantics, 4D geometric and motion priors, and action generation, all inside a Mixture-of-Transformers framework, abbreviated MoT. Each of those ingredients arrives already pretrained. The contribution is the arrangement that lets them serve one job, which is producing actions for a robot.

Inside that framework, a pretrained video expert and an action expert interact, under semantic guidance from a frozen VLM. A pretrained 4D foundation model injects geometric and motion priors through distillation that happens only during training. The authors further introduce Causal Imprint. It learns future-relevant scene changes from future supervision that is available only at training time, and it provides predictive representations directly to the action expert. At inference, no future video has to be rolled out. Both of these pieces share a pattern. The 4D priors and the future supervision are used while training and are not needed afterwards, so the model learns from them without having to carry them when it acts.

Joint training at large scale needs data to match. The authors construct a heterogeneous corpus of robot demonstrations, UMI data, egocentric human demonstrations, and Ego2Robot data, curated and aligned under a common state-action representation. The sources differ in kind: some record robots, some record humans from their own point of view, and aligning them under one shared representation is what lets a single model learn from all of them at once. The resulting corpus contains over twenty thousand hours of processed training data. To the authors' knowledge, that makes it the largest open-source corpus of its kind.

InternW0 Delta is pretrained on that corpus. The authors report that it outperforms prior methods across simulation benchmarks, and that it shows strong performance on real-robot platforms as well. The claim is made on both fronts, simulated and physical, rather than resting on a single kind of test.

The release plan is broad. The authors say they will open source the training code, the model weights, the infrastructure, the data-processing pipeline, and the processed data, where licenses permit. Underneath the architecture, the paper's move is integration: priors from video, semantics, geometry and motion brought under one framework, and a corpus large enough to train them together.