Robotics Has a 95% Data Gap. World Models Multiply Data — They Don't Create It.

The global stock of high-quality embodied interaction data, which consists of real robots performing real tasks recorded sufficiently for training, currently stands at approximately 500,000 hours as of early 2026. However, estimates suggest that a general-purpose embodied foundation model would require 10 million hours or more of such data.

This represents a shortfall of over 95%, unlike text data which can be easily scraped. The field's solution to this challenge is to use world models as data engines, training a generative model of environment dynamics and then using it to synthesize the trajectories that cannot be collected from real data. Systems like GigaWorld-0 render texture-varied scenes, novel viewpoints, and ego-centric translations, effectively inflating a small real dataset into a larger training corpus for vision-language-action (VLA) models, resulting in reported scale gains of 10-100x.

However, this approach also inadvertently recreates the failure modes encountered in language model land. As the synthetic-data loop progresses, distributional tails collapse first, leading to the thinning out of rare events, unusual lighting, and weird object geometries over generations. While synthetic data is excellent at interpolation, filling in the space between real data points, it is structurally incapable of extrapolation into physics it has never observed.

World models trained on rigid-body manipulation, for example, can generate confident, plausible, entirely wrong videos of cloth folds. The dynamics in these synthetically-generated videos are fiction, and training a VLA model on such fiction can lead to policies that fail on contact. The practical question, therefore, is not "real or synthetic," but "which slices of the distribution must be real, and how do you know when a synthetic slice has drifted?"

Three rules have emerged from working with teams building perception and action datasets. Firstly, real data anchors the contact boundary, where robot-robot or robot-world interactions take place, such as grasp initiation, slip, deformation, and force feedback. Synthetic augmentation is suitable for backgrounds, lighting, camera pose, and distractor objects, but not for the physics of contact moments.

Secondly, failure data is more valuable than success data, and it is often overlooked. Most robot datasets consist of demonstration datasets where an expert teleoperates the task correctly multiple times. These policies are not exposed to recovery scenarios, as they never see the robot struggling or the grasp slipping. Deliberately collecting near-misses and recovery sequences, and annotating the transition point where things go wrong, significantly improves downstream robustness compared to adding another 10,000 clean demonstrations.

Thirdly, a contamination check must be performed between the world model and the evaluation set. If the same generative model produced both training trajectories and evaluation scenarios, the benchmark measures self-consistency rather than competence. This is akin to the benchmark contamination problem encountered with frontier models this year, which could reproduce verbatim details of supposedly held-out tasks.

To ensure that the evaluation set consists of scenarios that no generative model in the stack has ever seen, real-world evaluation episodes must be held out and recorded on hardware. Although this process is slow and expensive, it is essential for obtaining trustworthy results. From a practical perspective, embodied data quality is mostly an annotation problem, and annotation for embodied AI is considerably more challenging than for image classification tasks.

A single manipulation episode may require various forms of annotation, including temporal segmentation, 3D spatial labels, cross-modal alignment, and outcome and failure-mode labels. Annotators must possess domain expertise to accurately determine the success or failure of an action and the specific frame at which a grasp becomes unrecoverable.

This level of annotation expertise is crucial due to the high cost and potential for errors associated with automated labeling methods. SyncSoft.AI, for instance, has developed a model for working with multimodal annotation, reasoning, and human feedback data programs, particularly focusing on LiDAR and video data for autonomous driving applications.