GAP: Scalable Driving with Generative Aided Planner
Keywords: end-to-end autonomous driving;generative;planning;perception
Abstract: The primary challenge in end-to-end autonomous driving lines in how to establish robust environmental perception and representations. While most methods improve these capabilities by introducing auxiliary perception tasks, the process of obtaining precise large-scale annotations in this paradigm is both time-consuming and laborious, thereby limiting the scalability and practical application. To address this, we propose an architecture based on the Generative Aided Planner (GAP), which integrates scene generation and planning within a single framework. To compensate for the information loss in discrete image features, we design a dual-branch image encoder that fuses continuous and discrete features, improving the model's ability to recognize traffic lights. Through the scene generation task from input tokens, our approach learns the intrinsic dependencies between tokens and environments, which in turn benefits the planning task. It is important to note that the generative model is trained in a fully self-supervised manner, requiring no perception annotations. Our model is built upon GPT-2, which exhibits scaling laws similar to those observed in other GPTs: as we increase the model size and data size, the performance shows continuous and non-saturating improvements. Experiments show that among methods using the front view as input, our approach outperforms other methods that employ multiple perception supervision in the CARLA simulator. Our method is simple yet highly effective, offering a promising direction for scalable and practical deployment of autonomous vehicles in real-world settings.
Primary Area: applications to robotics, autonomy, planning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5751
Loading