M-Predictive Spliner: Enabling Spatiotemporal Multi-Opponent Overtaking for Autonomous Racing

Nadine Imholz, Maurice Brunner, Nicolas Baumann, Edoardo Ghignone, Michele Magno

TL;DR

This paper introduces M-Predictive Spliner (M-PSpliner), a novel algorithm that enables autonomous race cars to safely and efficiently overtake multiple opponents simultaneously. By combining advanced multi-opponent tracking with predictive trajectory modeling, M-PSpliner significantly improves overtaking success rates and safety compared to existing methods, even at high speeds.

What Problem Does This Paper Address?

Autonomous racing pushes the boundaries of robotic decision-making, especially in complex scenarios like head-to-head racing with multiple competitors. While previous approaches have either simplified the problem by considering only one opponent or ignored the crucial spatiotemporal aspect of opponent behavior, this paper tackles the challenge of unrestricted multi-opponent overtaking. The core problem is how to predict the future movements of multiple dynamic agents and plan a safe, efficient overtaking maneuver in real-time, all within the constraints of a high-speed racing environment. Existing methods often rely on heuristics or struggle with the computational demands of multi-agent prediction, making them unsuitable for this demanding application.

Methods

The M-PSpliner framework builds upon the single-opponent Predictive Spliner (PSpliner) by introducing two key innovations:

  • Multi-Opponent Tracking with Re-Identification (reID): The system utilizes a Kalman Filter (KF)-based tracker, adapted from the image-based Norfair tracker, to operate in a spatial domain. This tracker is enhanced with a reID capability, allowing it to maintain the identity of individual opponents even through occlusions or brief periods of being out of sight. This is crucial for distinguishing between multiple cars on the track.
  • Spatiotemporal Multi-Opponent Prediction: For each tracked opponent, Gaussian Process Regression (GPR) is employed to model their spatial and velocity trajectories. This allows for the prediction of future movements. The system then identifies the "Region of Collision" (RoC) based on the closest opponent's predicted trajectory. While multiple opponents are tracked, the planning focuses on a single RoC, balancing computational efficiency with safety. If other opponents obstruct the planned overtaking path within the RoC, their positions are considered spatially during the final trajectory optimization.

The overtaking maneuver itself is planned using a sequential quadratic programming optimization, ensuring smoothness, adherence to track boundaries, and maintaining clearance from all detected opponents.

Datasets

The paper does not explicitly mention specific datasets for training. Instead, it focuses on experimental validation using a physical 1:10 scale autonomous racing car equipped with commercial off-the-shelf components and open-source software. The evaluation methodology involves performing multiple overtakes in controlled scenarios with varying opponent behaviors.

Key Insights and Ideas

  • Re-Identification is Key for Multi-Agent Tracking: The ability to reliably re-identify opponents is fundamental to extending single-agent tracking to multi-agent scenarios, especially in dynamic and potentially occluded environments like racing.
  • Focusing on a Single RoC is Sufficient: Despite tracking multiple opponents, computing a single "Region of Collision" based on the closest opponent proves to be an effective and computationally feasible strategy for planning. This avoids the significant computational overhead of calculating multiple RoCs.
  • Predictive Modeling Outperforms Heuristics: By using GPR for trajectory prediction, M-PSpliner moves away from heuristic-based opponent modeling, leading to more robust and safer overtaking maneuvers.
  • Minimal Overhead for Multi-Opponent Overtaking: The results show that overtaking multiple opponents does not significantly increase lap times or computational load compared to single-opponent scenarios, highlighting the efficiency of the proposed approach.

Key Results

  • Improved Overtake Success Rate: M-PSpliner achieved an overtaking success rate of up to 91.65%, a significant improvement of 10.13 percentage points over the previous state-of-the-art PSpliner in single-opponent scenarios.
  • Comparable Speed: The maximum speed scaler (Smax) for successful overtakes remained high, indicating that the safety improvements did not come at the cost of speed.
  • Efficient Multi-Opponent Handling: In multi-opponent scenarios (1 or 2 opponents), the average lap time remained nearly identical to the single-opponent case. The cumulative time for two overtakes was only slightly longer than twice the time for a single overtake, demonstrating efficient handling of multiple agents.
  • Accurate Tracking: The multi-opponent tracker achieved a positional error of 0.12m and a velocity error of 0.54 m/s in single-opponent scenarios, matching and slightly outperforming the baseline. Even with two opponents, the positional RMSE increased by only 0.06m.

Limitations and Future Work

  • Opponent Behavior Assumption: The current implementation assumes non-cooperative opponents, meaning they do not adapt their behavior to the ego agent's overtaking maneuvers (except for the "reactive opponent" type).
  • Future Work: The authors suggest integrating trajectory computation that ensures full feasibility, such as with kinematic Model Predictive Control (MPC), to further enhance the robustness of the overtaking maneuvers.

Summary updated

Abstract

Unrestricted multi-agent racing presents a significant research challenge, requiring decision-making at the limits of a robot's operational capabilities. While previous approaches have either ignored spatiotemporal information in the decision-making process or been restricted to single-opponent scenarios, this work enables arbitrary multi-opponent head-to-head racing while considering the opponents' future intent. The proposed method employs a KF-based multi-opponent tracker to effectively perform opponent ReID by associating them across observations. Simultaneously, spatial and velocity GPR is performed on all observed opponent trajectories, providing predictive information to compute the overtaking maneuvers. This approach has been experimentally validated on a physical 1:10 scale autonomous racing car, achieving an overtaking success rate of up to 91.65% and demonstrating an average 10.13%-point improvement in safety at the same speed as the previous SotA. These results highlight its potential for high-performance autonomous racing.
BibTeX
@inproceedings{imholz2025m,
  author = {Nadine Imholz and Maurice Brunner and Nicolas Baumann and Edoardo Ghignone and Michele Magno},
  title = {{M-Predictive Spliner: Enabling Spatiotemporal Multi-Opponent Overtaking for Autonomous Racing}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2025},
  doi = {10.1109/iros60139.2025.11247102},
}
Figure