Roadside perception systems need to anticipate what will happen next in a fixed coordinate frame, not just from a moving vehicle's perspective. InfraWorld, a new paper on arXiv, tackles this as a 4D occupancy forecasting problem: predicting future semantic occupancy in infrastructure coordinates.
The authors frame roadside scenes as 'globally structured yet locally evolving.' That means the system must hold onto comprehensive scene-level context while also accounting for local changes over time. InfraWorld is described as a motion-aware world model, suggesting it explicitly incorporates motion information to generate these forecasts.
The abstract is brief, but the contribution appears to be a modeling choice that balances global structure with local dynamics for roadside settings. This could matter for applications like autonomous driving and traffic management, where infrastructure sensors must reason about future states of the world.