Introduction
Autoregressive models have emerged as a powerful paradigm for autonomous driving planning, capable of capturing the multi-modal nature of human driving behaviors. Despite their promise, these models encounter a fundamental challenge: the "stale worldview" problem. In standard end-to-end configurations, a model processes sensor data at time $t$ to predict a sequence of future actions. However, as the vehicle moves, the spatial relationship between the initial sensor "snapshot" and the environment changes, leading to spatio-temporal misalignment.
Methodology
To address these limitations, this work proposes a three-pronged architectural improvement:
1. Time-Invariant Spatial Alignment (TISA)
Predicting future world states in raw pixel or voxel space is computationally prohibitive for real-time planning. TISA sidesteps this by performing geometric transformations directly in the latent space. By utilizing learnable key-value pairs, TISA "realigns" environmental features from the current timestamp to the ego-vehicle's predicted future poses. This ensures that each step of the autoregressive plan is conditioned on a geometrically consistent representation of the world.
2. Kinematic Action Head
To ensure physical feasibility, the model avoids predicting raw spatial waypoints. Instead, it outputs discretized values for acceleration and yaw rate. These control inputs are processed through a kinematic bicycle model to generate smooth, driveable trajectories, preventing "teleportation" artifacts common in simpler models.
3. Multi-Objective Policy Refinement (MOPT)
While imitation learning provides a strong baseline, it often fails to capture the nuances of safety and comfort. The authors employ Direct Preference Optimization (DPO) to refine the policy. Unlike standard DPO which uses a single scalar reward, this multi-objective approach generates preference pairs based on specific metrics such as Time-to-Collision (TTC) and Drivable Area (DA) compliance. This allows for targeted improvements in safety without sacrificing general driving performance.
Experimental Results
The model was evaluated on the NAVSIM benchmark, which utilizes the OpenScene dataset to provide a comprehensive testing ground with 360-degree camera and LiDAR data.
Performance Highlights
- State-of-the-Art Performance: The model achieved a 89.8 PDMS score, the highest recorded for autoregressive planners on NAVSIM.
- Efficiency over Scale: Despite using a lightweight ResNet-34 backbone, the model outperformed massive Vision-Language-Action (VLA) models like AutoVLA (3B parameters), highlighting the importance of architectural alignment over raw parameter count.
- Ablation Gains: The TISA module contributed a 1.2-point improvement in PDMS, while the Multi-Objective DPO stage provided a significant 3.0-point boost over pure imitation learning.
Conclusion
By addressing the core issue of spatio-temporal misalignment through geometric latent alignment and refining driving policies with multi-objective preferences, this work sets a new standard for end-to-end autonomous planning. The results demonstrate that specialized architectural components like TISA can achieve superior performance compared to general-purpose large-scale models, offering a more efficient path toward safe and reliable autonomous navigation.
Summary updated