Trajectory Evaluation via Rollout-Free World Model for End-to-end Autonomous Driving
Trajectory Evaluation via Rollout-Free World Model for End-to-end Autonomous Driving
This research introduces RFWorld, a rollout-free world model designed for end-to-end trajectory evaluation in autonomous driving. The method constructs a shared scene memory and utilizes lightweight temporal readout with semantic supervision during training. At inference, candidate trajectories query this memory via sparse spatial readout to score future relevance without reconstructing full future scenes, thereby reducing computational cost by approximately the product of the planning horizon and the number of candidates while avoiding accumulated prediction errors from recursive rollouts. Experiments on NAVSIM and Bench2Drive indicate that RFWorld achieves state-of-the-art performance with a world representation better aligned to planning requirements.
The paper proposes RFWorld, a world model architecture that avoids recursive rollout inference by using sparse spatial readout from a shared scene memory to evaluate candidate trajectories for end-to-end autonomous driving, demonstrating improved planning alignment and reduced computation on NAVSIM and Bench2Drive benchmarks.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
Source：arXiv Robotics — research abstracts · arxiv.org