Predictive Spliner: Data-Driven Overtaking in Autonomous Racing Using Opponent Trajectory Prediction

TL;DR

Predictive Spliner addresses overtaking in autonomous racing by learning opponent trajectories with Gaussian Process regression to forecast future opponent behavior and dimensions the RoC to plan overtakes. The method combines Opponent Trajectory Regression, Collision Prediction, and an Optimization-based Spliner in Frenet coordinates to produce feasible, safe overtakes that align with a global racing line. Empirical results on a 1:10 scale platform show substantial improvements over state-of-the-art, including overtaking at up to 83.1% of the ego speed and an 84.5% success rate, with real-time computation around 8.4 ms. The approach is open-source and demonstrated in competitive F1TENTH settings, with potential extensions to multi-opponent and game-theoretic frameworks for broader applicability.

Abstract

Head-to-head racing against opponents is a challenging and emerging topic in the domain of autonomous racing. We propose Predictive Spliner, a data-driven overtaking planner that learns the behavior of opponents through Gaussian Process (GP) regression, which is then leveraged to compute viable overtaking maneuvers in future sections of the racing track. Experimentally validated on a 1:10 scale autonomous racing platform using Light Detection and Ranging (LiDAR) information to perceive the opponent, Predictive Spliner outperforms State-of-the-Art (SotA) algorithms by overtaking opponents at up to 83.1% of its own speed, being on average 8.4% faster than the previous best-performing method. Additionally, it achieves an average success rate of 84.5%, which is 47.6% higher than the previous best-performing method. The method maintains computational efficiency with a Central Processing Unit (CPU) load of 22.79% and a computation time of 8.4 ms, evaluated on a Commercial off-the-Shelf (CotS) Intel i7-1165G7, making it suitable for real-time robotic applications. These results highlight the potential of Predictive Spliner to enhance the performance and safety of autonomous racing vehicles. The code for Predictive Spliner is available at: https://github.com/ForzaETH/predictive-spliner.

Figures (6)

• Figure 1: Qualitative visualization of the proposed Predictive Spliner overtaking planner for autonomous racing. The method leverages future opponent knowledge (yellow), initially regressing the opponent's trajectory via a (Step 1). The , defined by $c_{start}$ and $c_{end}$, is predicted by forward propagating the opponent's current and future poses through the (Step 2). Finally, this knowledge is used to compute an overtaking trajectory in the future section of the track (purple), avoiding premature overtaking maneuvers and improving performance (Step 3).
• Figure 2: Visualisation of the regressed opponent trajectory on an example track, sampled from $\mathcal{GP}_{d,v_s}(s)$ in $d$ and $v_s$ in yellow with the standard deviations shaded. Opponent detections obs from the robotic perception module in blue. Ground truth opponent position and speed in red.
• Figure 3: Example visualization of $c_{start,end}$ collision prediction. Current $t_0$ and future $t_{s,e}$ ego-agent poses ($P_{ego,t_i}$) in bronze, opponent poses ($P_{opp,t_i}$) in orange. Forward-propagated poses in grey, collision start ($c_{start}$) and end ($c_{end}$) at time $t_{s,e}$ in purple. The is shaded in purple. In this case, the ego racing line and the opponent racing line both coincide with the optimal racing line.
• Figure 4: Exemplified trajectory at the starting instant of the overtake. The -based evasion trajectory, shown with purple dots, avoids the opponent within the shaded . Poses of both the ego agent and the opp are shown with progressively shaded hues, to identify the progressing timesteps. At the end of the maneuver, the evasion trajectory reconnects the reference racing line, shown in red.
• Figure 5: Overtake comparison corresponding to the results of Table \ref{['tab:phys_res']} from left to right: Predictive Spliner (ours), Splinerforzaeth, gbo, and Frenetfrenetplanner. Each subplot illustrates the trajectory followed by the ego agent (colored path) and the opponent agent (gray path) during an overtaking maneuver. The maximum speed scaling $\mathcal{S}_{max}$ for each method indicates the highest speed at which the ego agent can successfully overtake the opponent. In all settings, the opponent is tracking the optimal racing line at their respective maximum speed scaling $\mathcal{S}_{max}$. The Predictive Spliner shows the highest $\mathcal{S}_{max}$, demonstrating its superior capability to overtake at higher speeds compared to the other methods.