A research team has built a sensor fusion backbone that learns from expensive LiDAR hardware during training, then ships without it.
CRISP is a spatiotemporal backbone that fuses camera and radar inputs into a bird's-eye-view representation of the road ahead. During pretraining, it ingests historical camera images and radar sweeps and learns by predicting future LiDAR point clouds — a self-supervised signal that forces the model to build a coherent spatial understanding of the scene. At deployment, the LiDAR is gone; only cameras and radar remain. The researchers tested the approach on the nuScenes benchmark and report improvements across six downstream tasks: 3D object detection, tracking, online mapping, motion forecasting, occupancy prediction, and planning.
The significance here is less about any single benchmark number and more about the training recipe. LiDAR sensors cost thousands of dollars per unit and are widely seen as impractical for mass-market vehicles. If a model can absorb LiDAR's geometric precision as a training signal and then generalize to cheaper sensor stacks, the gap between research-grade and consumer-grade autonomy narrows. It also fits a broader pattern in AI where privileged supervision — data available at training time but not inference time — is used to cheaply improve cheaper models.
Teams at Wayve, Tesla, and academic labs have all pursued world-model pretraining for driving over the past two years, but most work focuses on camera-only or LiDAR-heavy pipelines. Camera-radar fusion sits in an awkward middle: radar is cheap and works in bad weather, but its sparse, noisy returns have historically been hard to learn from. CRISP's radar-specific attention mechanisms — including Doppler-cue injection into temporal propagation — are the less-heralded contribution worth watching.