One model for all 3D perception tasks, online or offline

Most 3D perception systems are built separately for either real-time or batch processing, with different models for different sensors and scales. UniT unifies these into one framework using a Group Autoregressive Transformer that treats sensor observations in configurable batches and predicts 3D point maps without fixed anchors or scales. By adjusting group size, the same model runs in real-time (single frames) or offline (multi-frame batches), with memory-efficient caching for long sequences and a scale-adaptive loss that learns absolute metric scale. Tested on ten benchmarks across online perception, offline reconstruction, multi-modal fusion, and long-horizon tasks, it reaches top performance on all.