Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

PRIX: Learning to Plan from Raw Pixels for End-to-End Autonomous Driving

Created by
  • Haebom

Author

Maciej K. Wozniak, Lianhang Liu, Yixi Cai, Patric Jensfelt

Outline

PRIX (Plan from Raw Pixels) is an efficient end-to-end architecture that predicts safe paths for autonomous driving using only camera data. It removes the dependency on expensive traditional LiDAR sensors and computationally intensive BEV feature representations, and leverages a visual feature extractor and a generative planning head to predict safe paths directly from raw pixel inputs. Its core component, the Context-aware Recalibration Transformer (CaRT), effectively enhances visual features at multiple levels to enable more robust planning. It achieves state-of-the-art performance on NavSim and nuScenes benchmarks, and is comparable to larger, multi-modal diffusion planners while being much more efficient in terms of inference speed and model size. This makes it a practical solution suitable for real-world deployments.

Takeaways, Limitations

•
Takeaways:
◦
Demonstrates end-to-end autonomous driving using only cameras, without relying on LiDAR.
◦
It has a much more efficient inference speed and smaller model size than existing models.
◦
Achieving state-of-the-art performance on NavSim and nuScenes benchmarks.
◦
It is released as open source, increasing the possibility of practical application.
•
Limitations:
◦
Additional performance validation in actual road environments is required.
◦
Further research may be needed on the generalization performance of the CaRT module.
◦
There is a possibility of performance degradation in extreme environments (bad weather, extreme lighting changes, etc.).
👍
Made with Slashpage