One LiDAR model now simulates, forecasts, and rebuilds worlds
LidarVerse puts LiDAR world generation, point-cloud forecasting, and scene resimulation in one real-time model, so researchers can switch tasks by changing the scene evidence they provide instead of optimizing each scene separately. It reads LiDAR sequences into a shared latent space, then uses a four-step causal model with temporal and spatial memory to sustain long rollouts. The paper reports 11 FPS on one H100 GPU and support for 32-, 64-, and 128-beam sensors, plus an unseen scanning pattern.