TLDR: A new framework called MAP (Map-Assisted Planning) significantly improves end-to-end autonomous driving by explicitly integrating online map features and the vehicle’s current status into trajectory planning. This novel approach, which includes a Plan-enhancing Online Mapping module, an Ego-status-guided Planning module, and a dynamic Weight Adapter, leads to substantial reductions in L2 displacement error and off-road rates. MAP achieved top ranking in a recent challenge, demonstrating its effectiveness in creating more robust and accurate autonomous driving systems without relying on complex, stacked modules.
End-to-end autonomous driving systems, which aim to integrate perception, prediction, and planning into a single framework, have gained considerable attention. However, a common challenge in many existing approaches is the underutilization of online mapping modules, leaving a significant potential for enhancing trajectory planning untapped.
A new framework called MAP (Map-Assisted Planning) addresses this limitation by proposing a novel way to explicitly integrate map features and the vehicle’s current status into the planning process. This approach aims to create a trajectory planner that is both robust and accurate.
Traditional autonomous driving systems often struggle because the outputs from various modules, like map segmentation, don’t directly guide the planning process. This can lead to situations where map information, which provides crucial spatial context such as road topology, drivable areas, and obstacle boundaries, isn’t fully leveraged. Without proper integration of this spatial context, planning models might rely too heavily on local dynamics, potentially leading to less reliable or short-sighted driving decisions.
MAP tackles this by introducing three key components:
Plan-enhancing Online Mapping (POM) Module
This module generates map-guided queries that are directly decodable into trajectories. It takes the detailed map features produced by the online mapping module and combines them with the vehicle’s current status to create a planning query that explicitly influences trajectory decoding.
Ego-status-guided Planning (EP) Module
Operating in parallel, this module focuses on the vehicle’s current status, including its velocity, acceleration, heading angle, and driving commands. It uses these real-time inputs to generate an ego-guided planning query, ensuring that the vehicle’s immediate state contributes to basic planning robustness.
Also Read:
- DGMap: A New Approach to Road Network Inference
- Advancing Autonomous Driving with Large Foundation Models for Trajectory Prediction
Weight Adapter
This intelligent component dynamically fuses the outputs from both the POM and EP modules. Based on the vehicle’s current status, it learns to assign a weight to each query, creating a combined representation that is then decoded into the final driving trajectory. This adaptive fusion ensures a harmonious cooperation between map-based context and real-time ego status.
Experiments conducted on the DAIR-V2X-seq-SPD dataset demonstrated significant improvements. MAP achieved a 16.6% reduction in L2 displacement error, a 56.2% reduction in off-road rate, and a 44.5% improvement in overall score compared to the UniV2X baseline, even without additional post-processing. Furthermore, the framework secured the top ranking in Track 2 of the End-to-End Autonomous Driving through V2X Cooperation Challenge at the MEIS Workshop @CVPR2025, outperforming the second-best model by 39.5% in overall score.
These results underscore the effectiveness of explicitly leveraging semantic map features in planning and suggest new directions for designing more robust and efficient end-to-end autonomous driving systems. The research paper can be found here: MAP: End-to-End Autonomous Driving with Map-Assisted Planning.
