PRIX (Plan from Raw Pixels) is an efficient end-to-end architecture that predicts safe paths for autonomous driving using only camera data. It eliminates the dependency on existing large-scale models, expensive LiDAR sensors, and computationally intensive BEV (Bird's Eye View) feature representations, and utilizes a generative planning head and visual feature extractor that predict paths directly from raw pixel inputs. The core component, Context-aware Recalibration Transformer (CaRT), effectively enhances multi-level visual features to enable more robust planning. It achieves state-of-the-art performance on NavSim and nuScenes benchmarks, and is comparable to large-scale multimodal diffusion planning models while being much more efficient in terms of inference speed and model size. Therefore, it is evaluated as a practical solution suitable for real-world deployment. The source code will be released.
