SDS++: Online Situation-Aware Drivable Space Estimation for Automated Driving

TL;DR

SDS++ addresses the challenge of sustaining accurate drivable space in dynamic driving contexts by replacing SDS's graph-based output with a lightweight artificial potential field (APF) representation and introducing an implicit line factor for SLAM. The approach enables standardized output compatible with planners and provides a robust, situation-aware drivability estimation by combining bounding-box and implicit-line drivability terms with domain knowledge. Key contributions include a novel implicit function factor for line features, an APF-based object drivability model with sigmoid compositions, and validation using real vehicle data in conjunction with an MPC-based planner. The results show improved localization robustness to sensor noise, more accurate object geometry handling, and feasible real-time performance, highlighting SDS++’s practical impact for online planning in challenging environments.

Abstract

Autonomous Vehicles (AVs) need an accurate and up-to-date representation of the environment for safe navigation. Traditional methods, which often rely on detailed environmental representations constructed offline, struggle in dynamically changing environments or when dealing with outdated maps. Consequently, there is a pressing need for real-time solutions that can integrate diverse data sources and adapt to the current situation. An existing framework that addresses these challenges is SDS (situation-aware drivable space). However, SDS faces several limitations, including its use of a non-standard output representation, its choice of encoding objects as points, restricting representation of more complex geometries like road lanes, and the fact that its methodology has been validated only with simulated or heavily post-processed data. This work builds upon SDS and introduces SDS++, designed to overcome SDS's shortcomings while preserving its benefits. SDS++ has been rigorously validated not only in simulations but also with unrefined vehicle data, and it is integrated with a model predictive control (MPC)-based planner to verify its advantages for the planning task. The results demonstrate that SDS++ significantly enhances trajectory planning capabilities, providing increased robustness against localization noise, and enabling the planning of trajectories that adapt to the current driving context.

Figures (13)

• Figure 1: Examples of challenging driving situations for an AV.
• Figure 2: Typical architecture of an AV (simplified), featuring our proposed online situation-aware drivable space (SDS++) estimation. Raw input data is processed to build a world model, which is then used for planning and executing the best course of action. Our SDS++ enables the construction of a situation-aware drivable space and ego localization within it, allowing the AV to navigate complex scenarios that traditional environment representations cannot handle.
• Figure 3: Example surface interpretation of implicit line function fitting using 3L, with $d_{3L}=1$ and $\|\nabla_{des}\|=5$. All line points are close to the fitted surface.
• Figure 4: Sigmoid function with varying $\beta$ and $\alpha=1$.
• Figure 5: Bbox edges and drivability example with $\beta=5$.