QuayPoints: A Reasoning Framework to Bridge the Information Gap Between Global and Local Planning in Autonomous Racing

TL;DR

The paper tackles the information gap between global time-optimal racelines and online local planning in autonomous racing. It introduces QuayPoints by solving multiple minimum-time racelines while sweeping lateral bounds, and identifying regions where the normalized lateral position is invariant across solutions as time-critical anchors, using the criterion $\sigma(s) < \tau$ with $\tau = 0.10$. QuayPoints provide a compact, planner-agnostic channel to convey global time-optimality to the local planner with minimal online overhead, demonstrated by integration into a multilayer graph-based local planner. Offline perturbation studies show that deviations at QuayPoints incur larger lap-time penalties (averaging $1.26\%$) than non-QuayPoint regions (about $0.50\%$), and a QuayPoint-aware planner improves overtaking performance up to $75\%$ of the ego speed across four tracks, validating the practical utility of the approach. The framework is general, extendable to other optimality criteria and data-driven extractions, and provides a scalable path to close the global-to-local information gap in autonomous racing and related domains, with potential extensions to graded QuayPoint salience and learning-based predictors. $T_f = \int_0^1 \frac{1}{\dot{s}} ds$ and $\sigma(s) < 0.10$ are used to characterize time-optimal trajectories and QuayPoints, respectively.

Abstract

Autonomous racing requires tight integration between perception, planning and control to minimize latency as well as timely decision making. A standard autonomy pipeline comprising a global planner, local planner, and controller loses information as the higher-level racing context is sequentially propagated downstream into specific task-oriented context. In particular, the global planner's understanding of optimality is typically reduced to a sparse set of waypoints, leaving the local planner to make reactive decisions with limited context. This paper investigates whether additional global insights, specifically time-optimality information, can be meaningfully passed to the local planner to improve downstream decisions. We introduce a framework that preserves essential global knowledge and conveys it to the local planner through QuayPoints regions where deviations from the optimal raceline result in significant compromises to optimality. QuayPoints enable local planners to make more informed global decisions when deviating from the raceline, such as during strategic overtaking. To demonstrate this, we integrate QuayPoints into an existing planner and show that it consistently overtakes opponents traveling at up to 75% of the ego vehicle's speed across four distinct race tracks.

Figures (7)

• Figure 1: Track representation. The arrow specifies the direction on intended lap
• Figure 2: QuayPoints : time-critical regions revealed by invariance of lateral position across perturbed, time-optimal trajectories. QuayPoints are calculated by first solving the minimum-time trajectory multiple times while sweeping the allowable lateral position limits $(\lambda_{\min}(s),\,\lambda_{\max}(s))$ along the track (a) . For each solution, record the raceline’s lateral position ($lambda(s)\in[0,1]$), measured across the track from outer edge (0) to inner edge (1) at path parameter ($s$). (b) At each path coordinate, stack the solutions and normalize each raceline’s lateral position to a zero-to-one scale using the local lower and upper limits and measure the across-solution spread (standard deviation) of this normalized position. (c) Shows the std. dev $\sigma(s)$ computed from a sweep of 55 racelines for Nürburgring circuit. Thresholding ($\sigma(s)<\tau$) ($\tau=0.10$) yields QuayPoints that can guide the local planner better.
• Figure 3: Average time loss if the car deviates from the raceline at a QPt vs at non-QPt. On average a single deviation at QPt results in a laptime that is 1.26% higher to a single deviation at non-QPt which results in a 0.5% higher laptime
• Figure 4: All deviated paths converge at the QuayPoint region (dark red), regardless of where the raceline is blocked. Obstacles of a color only exist for the raceline of the corresponding color. The white path is the raceline without any obstacles.
• Figure 5: Overtake attempts by kp aware method (Top) and without Kp (Bottom).