Hello Inc.'s robotaxi business has released HelloWorld, a generative driving world model designed to build driving data on demand and to test driving policies interactively. The team announced the model on October 12 and has published a paper on arXiv along with a project page.
HelloWorld brings together scene control, continuous generation, multi-sensor observation and efficient inference. Engineers can edit vehicles, pedestrians and their trajectories directly, and the model generates matching multi-view camera video together with lidar observations. Examples include inserting a pedestrian crossing the road, changing the route of the car ahead, or switching between rain, snow, day and night.
The model also accepts plain-language descriptions as a second way to build scenes. A user describes the situation they want, and an application pipeline turns it into the conditions the model needs. Construction zones, dense traffic, pedestrians crossing, fog and different lighting can all be used to steer what is generated.
The goal is to fill gaps in real-world data. Long-tail situations, such as rare hazards or unusual combinations of weather and traffic, are hard to capture enough of through road collection alone. Generating them on request lets developers create targeted training data and run counterfactual tests: replaying the same moment with one element changed to see how a driving policy responds.
According to the paper, HelloWorld has about 2 billion parameters, produces views for seven cameras with conditional lidar output, and uses causal few-step generation. That design is meant to keep inference fast enough for repeated, interactive simulation rather than one-off clip generation.
World models have become a common tool across autonomous driving and embodied AI, where teams use them to cut the cost of collecting and labeling edge cases. HelloWorld is the robotaxi team's first public research release of this kind. The company has not said when or how the model will be used in its vehicle testing programs.
SEE ALSO: Xiaomi Open-Sources Robotics-U0 Embodied World Model and Training Stack · Amap Releases ABot-Earth 0.7 3D-Native Urban World Model