Abstract
While supervised learning is widely used for perception modules in conventional autonomous driving solutions, scalability is hindered by the huge amount of data labeling needed. In contrast, while end-to-end architectures do not require labeled data and are potentially more scalable, interpretability is sacrificed. We introduce a novel architecture that is trained in a fully self-supervised fashion for simultaneous multi-step prediction of space-time cost map and road dynamics. Our solution replaces the manually designed cost function for motion planning with a learned high dimensional cost map that is naturally interpretable and allows diverse contextual information to be integrated without manual data labeling. Experiments on real world driving data show that our solution leads to lower number of collisions and road violations in long planning horizons in comparison to baselines, demonstrating the feasibility of fully self-supervised prediction without sacrificing either scalability or interpretability.
Introduction
Traditional autonomous driving (AD) systems typically rely on a modular pipeline where perception, prediction, and planning are separate stages. While this provides interpretability, the planning stage often depends on manually designed cost functions that are difficult to tune for the vast complexity of real-world driving. Conversely, end-to-end deep learning models offer scalability by learning directly from data but often lack the transparency required for safety-critical applications. This research proposes a middle ground: a self-supervised framework that predicts interpretable intermediate representations—Occupancy Grid Maps (OGMs) and Cost Maps (CMs)—to guide motion planning without requiring manual labeling.
Methodology
The proposed architecture processes LiDAR point clouds and HD map information to simultaneously predict future road dynamics and driving costs. The system outputs two primary components: Occupancy Grid Maps (OGMs), which predict the future spatial occupancy of dynamic obstacles, and Cost Maps (CMs), which provide a high-dimensional representation of the desirability of driving in specific locations.
Self-Supervised Learning Framework
A core challenge in learning cost maps is the lack of ground truth. The authors address this through a combination of environmental priors and an Auxiliary Imitation Task. Environmental priors penalize the model for assigning low costs to known obstacles or non-drivable regions. The auxiliary task uses an intention network to predict the expert driver's trajectory, effectively using imitation learning as a guide to shape the cost map in areas where no obstacles are present.
The researchers explored two variants: the Recurrent Cost Map Estimator (RCME), which leverages temporal dependencies, and the Multi-Step Cost Map Estimator (MSCME), which predicts the entire horizon in a single pass. To improve dynamic object handling, the OGM predictor utilizes "Difference Learning" to focus on frame-to-frame changes.
Experimental Results
The framework was evaluated using the Argoverse dataset, with raw sensor data converted into Bird’s-Eye View (BEV) grids. Performance was measured using the minimum Average Displacement Error (minADE), Collision Rate (CR), and Road Violation (RV) rate.
As shown in the results table, the proposed RCME models significantly outperform standard Behavior Cloning (BC) and rule-based systems, especially in long-range planning. For a 3-second horizon ($T=3$), the RCME.b.1 variant achieved a collision rate of 0.78%, compared to 7.11% for BC-MLP and 4.93% for RuleCM. Furthermore, the road violation rate remained exceptionally low (0.01%), indicating successful integration of HD map constraints.
Conclusion
This work demonstrates that self-supervised learning can produce interpretable and high-performing cost maps for autonomous driving. By leveraging auxiliary imitation tasks and environmental priors, the model bridges the gap between modular interpretability and end-to-end scalability. Future work will focus on integrating more sophisticated differentiable planners and explicitly modeling multi-agent interactions to further enhance safety in dense traffic scenarios.
Summary updated