EVO-MPCC: Enhanced Velocity Optimization with Learning-Based Auto-Tuning for Real-Time Vehicle Trajectory Planning

Feasible and Aggressive Cornering Trajectories

Exposure time of 4s

Exposure time of 2.5s

EVO-MPCC Enables Planning Feasible Cornering Trajectories on Mixed Racetracks. The figure compares the trajectories of the proposed Enhanced Velocity Optimization MPCC (EVO-MPCC) and other optimal control problem (OCP)-based planners, in high-curvature corners with the RoboRacer vehicle. The driven trajectories are visualized through long-exposure photography, illustrating that EVO-MPCC outperforms the benchmarks by achieving higher progress along the racetrack.

Abstract

Model Predictive Contouring Control (MPCC) is widely adopted in autonomous racing for its effectiveness in modeling racing progress; however, existing versions struggle to balance path progress and cornering feasibility on racetracks with rapidly varying curvature and rely heavily on manual parameter tuning, which may lead to unsafe behaviors. To address these challenges, this paper proposes an Enhanced Velocity Optimization Model Predictive Contouring Control (EVO-MPCC) framework that explicitly incorporates a reference velocity profile (RVP) into the MPCC objective and employs Bayesian optimization for automatic parameter tuning. By performing continuous velocity optimization along the racetrack, EVO-MPCC enables feasible cornering and high-performance racing even under a limited prediction horizon. Based on the RVP, two complementary formulations are developed: EVO-RVT, which performs Reference Velocity Tracking (RVT) for high-performance racing in obstacle-free scenarios, and EVO-TVC, which introduces an RVP-based Terminal Velocity Cost (TVC) to enable flexible, collision-free, and time-efficient overtaking. Additionally, a lap-time minimization (LTM) objective is formulated for Bayesian optimization to tune high-performance planner parameters automatically. The proposed framework is validated on the RoboRacer platform through extensive simulation and real-world experiments. Results show that EVO-TVC achieves a 9.81% relative lap-time advantage over hierarchical overtaking approaches. At the same time, EVO-RVT reaches 98.23% of the reference projected velocity with a mean computation time of 6.41 ms, demonstrating consistent high-speed performance over multiple consecutive laps, and a 50.42% improvement in energy efficiency relative to offline-planned trajectories.

Proposed Framework for LTM-Based BO Auto-Tuning of the Parameters of a new EVO-MPCC Trajectory Planner. The black arrows represent the evaluation process of the current parameter, while the light brown arrows denote the next parameter selected by the acquisition function for evaluation. The closed-loop training terminates at the N_BO iteration.

Motivation

Key findings: Conventional MPCC parameter tuning is challenging on mixed racetracks.

Issues and conflicts of short-prediction-horizon MPCC: From the formulation of short-horizon MPCC, mixed racetracks inherently impose mutually conflicting requirements on a single set of controller parameters. On high-curvature corners, feasible and safe trajectory planning requires a large penalty on the lag error together with a relatively small progression weight to ensure cornering feasibility. In contrast, high-performance racing on long straight segments demands aggressive acceleration and rapid path progression, which are typically achieved by reducing the lag error weight and increasing the progression weight. These fundamentally contradictory parameter requirements must be satisfied simultaneously across heterogeneous racetrack geometries, forcing the MPCC to adopt a conservative parameterization dominated by the most restrictive scenarios—sharp corners. Consequently, acceleration capability on straight segments is unnecessarily limited, leading to suboptimal exploitation of straight-line performance and a degradation in overall lap time.

Numerical Simulation Results for Racing and Obstacle Avoidance

Racetracks Employed in Simulation and Real-world Experiments. Each track includes its offline-calculated raceline (minimum-curvature path) and ggv-constrained corresponding reference velocity profiles (RVPs) in m/s. Track1 and Track2 are self-constructed, while Track3 is obtained from the official competition racetrack of the 18th Roboracer Autonomous Grand Prix.

Our results indicate that the best lap times achieved by EVO-RVT through BO training outperform the offline trajectories on racetracks with consecutive corners. This improvement is realized by shortening the racing trajectory. Furthermore, the trained parameters exhibit the ability to generalize across various racetracks. In addition, the proposed LTM-based objective enables consistently strong BO training performance across different racetracks, demonstrating its practicality for track-specific parameter tuning.

EVO-MPCC with TVC (EVO-TVC) Enables Collision-free Overtaking near the Limits. The Terminal Velocity Cost (TVC) provides increased replanning flexibility, allowing the vehicle to avoid obstacles without excessively slowing down. In contrast, RVT greatly reduces its speed when approaching the obstacles, as shown in (a) and (b).

Real-World Validation of Feasibility and Aggressiveness

EVO-MPCC with RVT (EVO-RVT) Maintains Smooth And Feasible Racing Trajectories In Consecutive Corners And Sharp Turns, Achieving Aggressive Yet Feasible Cornering Without Running Close To The Track Boundaries. Its trajectories remain consistently close to the offline reference and exhibit more uniform cornering behavior, whereas several baselines show either conservative detours or increased boundary proximity in sharp corners. These observations highlight the high feasibility and racing performance of EVO-RVT in complex cornering scenarios.

Comparison of computation efficiency and lap time repeatability.

Computational Efficiency (a) and Repeatability (b) of EVO-RVT Over Multiple Racing Laps, with the Real-World RoboRacer Vehicle. In the absence of disturbances in the localization module (TG0), our EVO-RVT maintains highly repeatable lap times over 10 consecutive laps.

First Viewpoint replay

Third Viewpoint

replay

EVO-MPCC has been successfully deployed on the real-world RoboRacer platform, demonstrating both feasible and aggressive racing. It secured 4th place in the Roboracer competition at IV2024.

Conclusion

This paper presents an Enhanced Velocity Optimization MPCC (EVO-MPCC) framework for high-performance autonomous racing. By explicitly integrating velocity optimization into MPCC, the proposed approach effectively balances aggressive racing behavior with stable cornering performance under limited prediction horizons. An LTM-based Bayesian Optimization (BO) module is further introduced to enable efficient and automated parameter tuning. Real-world experiments demonstrate that EVO-RVT consistently achieves a mean projected velocity of 98.23% of the reference performance limit, while maintaining real-time feasibility with an average computation time of only 6.41 ms. In addition, EVO-RVT reliably completes multiple consecutive laps and improves onboard energy efficiency by more than 50% compared to offline-planned trajectories, highlighting its robustness and efficiency in real-world racing scenarios.
The proposed EVO-MPCC framework comprises two complementary formulations targeting different racing tasks. EVO-RVT focuses on high-performance racing in obstacle-free scenarios by continuously optimizing vehicle velocity along the racetrack using a reference velocity profile, allowing the planner to accurately identify appropriate braking points and generate feasible cornering trajectories even with a short prediction horizon. EVO-TVC extends this framework to overtaking scenarios by introducing reference-velocity-based terminal velocity cost together with adaptively weighted collision-avoidance costs, thereby preserving sufficient maneuvering flexibility while ensuring safety. Experimental results show that EVO-TVC achieves a 9.81% reduction in lap time compared to hierarchical overtaking planners, demonstrating the advantages of unified planning-and-control optimization for time-efficient and collision-free overtaking.

BibTeX


      To be updated soon.

A brief conference paper only employs reference velocity tracking for racing: A Data-Driven Aggressive Autonomous Racing Framework Utilizing Local Trajectory Planning with Velocity Prediction
A simpler approach uses the racetrack's offline curvature to optimize the velocity profile of the local trajectory: Reduce Lap Time for Autonomous Racing with Curvature-Integrated MPCC Local Trajectory Planning Method