MAP2: Model- and Acceleration-Based Pursuit with MPC and Gaussian Process Residual Learning for Autonomous Racing

arXiv:2610.12196 · cs.RO · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "MAP2: Model- and Acceleration-Based Pursuit with MPC and Gaussian Process Residual Learning for Autonomous Racing".

Dev: The gist Autonomous racing requires accurate trajectory tracking near handling limits while maintaining low computational latency,

Rosa: First, who's behind it and why it matters.

Paper summary: Rosa: So, we're looking at this paper called "MAP2: Model- and Acceleration-Based Pursuit with MPC and Gaussian Process Residual Learning for Autonomous Racing." Basically, they’re tackling the problem of needing to track a car really accurately when it’s pushing the limits. Geometric controllers are fast but they break down when you're near those limits because the Ackermann steering geometry isn't accurate anymore.

Dev: Yeah, and what MAP2 does is try to keep that simplicity while adding more physics in by using tire dynamics, but they realized that even with those tire models, there’s a gap between what the model predicts and what the real car actually does.

Rosa: They combine a curvature-based kinematic MPC with this Sparse Gaussian Process residual correction to fix that mismatch. The core idea is using an MPC to plan out the control inputs over time, and then using a learned correction to adjust those steering commands based on how much the real car deviates from what the model expects.

Dev: So, it’s not just one thing; it’s a two-part system: first, planning with MPC using a kinematic model like Frenet frame dynamics, and second, learning from real data with an SGP to make sure the steering commands are actually right when they need to be. This whole approach is aimed at getting better tracking performance while keeping the computation low enough for real-time use.

Taro: From my side, I'm interested in how this system handles things when the world gets messy, like when you hit model mismatches or tire parameters change unexpectedly. The paper claims this method provides robustness to those modeling errors by learning a residual correction from real-world driving data collected on an F1TENTH vehicle.

Rosa: Exactly, and that’s where the SGP comes in. Instead of relying solely on a fixed model, they feed the SGP input features like velocity, curvature, and yaw rate to predict that steering correction term, c delta k.

Dev: The math shows they use this correction to adjust the nominal steering command before it gets limited. That means instead of just following what the nominal model says for your steering angle, you’re adding this learned adjustment to get the actual command, delta pre k = delta nom k + c delta k.

Taro: It sounds like they are specifically addressing that limitation where pure pursuit controllers can't guarantee the vehicle stays within the limits, especially at high speeds where those geometric assumptions fail.

Rosa: That’s right; the paper shows that this combination of MPC for prediction and SGP for correction leads to better lap times compared to existing methods like MAP and PP, even when things get tough. This whole setup is designed for accurate trajectory tracking near the handling limits with low latency.

Conclusion: Rosa: So, wrapping up this discussion on "MAP2," the main thing is that they took a geometric idea, like Model- and Acceleration-based Pursuit, made it more predictive with MPC, and then patched it up with a Sparse Gaussian Process to handle real world errors. It’s about making these complex maneuvers reliable in high-performance racing.

Dev: The authors are Huang, Hu, Ghignone, Baumann, Wang, Su, Xie Magno. They put this implementation out there for reproducibility at https: //anonymous.4open <ref:2610.12196#pg1>.science/r/MAP2-1F88/ <ref:2610.12196#pg1>. It’s a complete controller setup that works on a one:ten scale F1TENTH platform <ref:2610.12196#pg1>.

Rosa: What this means for us on the show is that we have something computationally lightweight enough to run in real time, about nine milliseconds per cycle, which is crucial for actual autonomous racing systems <ref:2610.12196#pg1>. It’s not just theoretical work; it’s a controller that shows tangible improvements in tracking errors and lap times compared to older methods.

Dev: And the results show that when you test it at high speeds, like seventy-five percent, MAP2 cuts the average lateral tracking error by thirty-seven point nine nine percent compared to MAP, and even better against the classical PP controller. That level of improvement shows how much accounting for model mismatch can help maintain performance under stress.

Taro: For someone who just listens to the show, this paper suggests that for autonomous systems, you don't just need a perfect model; you need a system that can actively learn and correct its own mistakes in real time when things go off script.

Rosa: That’s the big picture—it’s not about achieving perfect simulation; it’s about building something robust enough to handle the messy reality of driving at high speed. This MAP2 framework shows how combining prediction, model-based pursuit, and residual learning can result in a control system that's both fast and adaptable.

Dev: We look forward to seeing how this kind of MPC approach scales up to more complex vehicle dynamics where the tire forces become even less predictable. That’s the next step for this research.

Weizhan Huang, Cheng Hu, Edoardo Ghignone, Nicolas Baumann, Yiqin Wang, Hongye Su, Lei Xie

Zhejiang University

cs.RO

Submitted: 2026-10-08

Updated: 2026-10-08

The gist: The gist Autonomous racing requires accurate trajectory tracking near handling limits while maintaining low computational latency, and MAP2 addresses this by combining curvature-based kinematic MPC

Key concepts

Kinematic MPC
This optimization technique predicts future control inputs over a short time horizon. In MAP2, it is used to calculate optimal lateral acceleration requests based on vehicle kinematics, forming the core predictive component of the control strategy.
Sparse Gaussian Process (SGP)
SGP is a machine learning method used here to learn and correct errors. It takes real-world driving data as input and learns a 'residual'—the difference between what the model predicts and what actually happens—to provide an accurate correction for steering commands.
Frenet-frame Kinematic MPC
This is a specific way to formulate the MPC problem. It uses a simplified kinematic model based on vehicle motion (like curvature) instead of complex tire force dynamics, making the optimization computationally efficient while still providing good predictive control.
Model Mismatch Correction
This addresses the problem where the steering commands derived from a pre-calculated lookup table (LUT) don't perfectly match actual vehicle behavior. MAP2 uses SGP to learn and apply a correction term ($Δcδk) to adjust these nominal commands, ensuring better performance when tire parameters change.

Terminology

Summary

The gist Autonomous racing requires accurate trajectory tracking near handling limits while maintaining low computational latency, and MAP2 addresses this by combining curvature-based kinematic MPC with Sparse Gaussian Process residual learning to enhance performance compared to existing methods.

How it works

  1. The proposed algorithm uses MPC to optimize kinematic control inputs over a prediction horizon and maps them to steering commands through a tire dynamics model augmented with SGP residual correction

  2. A Frenet-frame kinematic MPC is introduced to replace the geometric look-ahead based lateral acceleration calculation in MAP

  3. The SGP residual model is learned from real-vehicle data to compensate for this mismatch by correcting the LUTbased steering commands

Key Components of MAP2

** MPC-Based Predictive Lateral-Acceleration Generation:**

The Frenet kinematic model is used as a simple representation of the vehicle dynamics, excluding the tire forces. The optimization problem is formulated as min uk+ik,ϵ N X−1 i=0 [qn · n 2 k+ik + qeψ · e 2 ψ,k+ik +∥uk+ik − u ref k+ik∥ 2 R] + pn · n 2 k+Nk + peψ · e 2 ψ,k+Nk + ρ · ϵ 2(6). The first stage of the input for the MPC solution is used to calculate the optimal lateral acceleration request ac: a∗c = q (v∗ x) 2 + (v∗ y) 2 · Ψ˙ ∗(12).

** Model- and Acceleration-based Pursuit:**

The basic principle of MAP is using the vehicle’s lateral dynamics to establish a new mapping relationship. This mapping uses the dynamic motions of the lateral velocity vy and the yaw rate Ψ˙ expressed as v˙y = 1/m [Fy,r + Fy,f · cos δ + Fx,f · sin δ] − vx · Ψ˙ (16) and Ψ =¨ 1/Iz [−lrFy,r + lf (Fy,f · cos δ + Fx,f · sin δ)] (17). The nominal steering command is obtained by interpolating between the closest elements in the table.

** Sparse Gaussian Process Residual Correction:**

Due to the mismatch between the LUT and real vehicle dynamics, MAP2 learns a steering residual from realworld driving data collected on the F1TENTH vehicle. The SGP input feature vector is zk = [vx,k, rk, ax,k, ay,k, δprev,k] T, (24). The trained sparse GP predicts the steering-residual correction ∆cδk and the corrected steering command before limiting is computed as δ pre k = δnom k + ∆cδk (25).

Experimental Validation and Results

The proposed controller was experimentally validated on a 1:10-scale F1TENTH platform on two different tracks and over multiple speed settings. The experiments demonstrate reduced lateral tracking errors and lap times at high speeds relative to MAP and PP, with additional improvements from SGP residual correction.

** Performance Comparison:**

At the highest speed of 75%, compared with the baseline MAP controller, MAP2 reduces the average lateral tracking error by 37.99%, the maximum lateral tracking error by 57.48%, and the lap time by 1.57%. Compared with the classical PP controller, according to Table V, the average and maximum lateral tracking errors are reduced by 44.65% and 47.81%, respectively, together with a 3.27% reduction in lap time.

** Robustness to Model Mismatch:**

The cross-tire experiments further show the importance of residual correction when the tire parameters used to generate the offline LUT no longer accurately represent the actual vehicle dynamics. MAP2 substantially improves tracking performance over the original MAP controller under this model mismatch, reducing the average and maximum lateral tracking errors by 23.89% and 36.89% at the velocity of 70%, respectively, and at the velocity of 75%, MAP2 finished 10 laps with good lateral error control, where MAP crashed into the wall.

Conclusion

In this paper, we proposed MAP2, a computationally lightweight and robust lateral controller for high-performance autonomous racing. The proposed framework combines a kinematic MPC for predictive lateral-acceleration generation, a model-based LUT for steering computation, and an SGP for residual correction. This combination improves predictive tracking capability and compensates for model mismatch while maintaining low computational latency. The complete MAP2 control cycle requires approximately 9 ms on the onboard computing platform, demonstrating its suitability for real-time autonomous racing on computationally constrained systems.

--- Page 7 ---

TABLE IV: Controller performance on the physical system in map I Velocity scale (%) Controller Avg. Lateral error (m) Max. Lateral error (m) Lap time (s) Avg. Std. Avg. Std. Fastest Avg. Std.

--- Page 8 ---

TABLE V: Controller performance on the physical system in map II Velocity scale (%) Controller Avg. Lateral error (m) Max. Lateral error (m) Lap time (s) Avg. Std. Avg. Std.

--- Page 7 ---

Figure 7: Effect of SGP on lateral-acceleration tracking for a representative MAP2 segment.

--- Page 8 ---

Figure 8: The comparison of the controllers when tires and tire parameters mismatch (MAP in 75% velocity crashed into the wall and did not finish).

--- Page 5 ---

TABLE I: Comparison of representative autonomous racing controllers Controller Compute High-speed tracking Model sensitivity Low Limited Low Improved High

--- Page 3 ---

Figure 3: 1:10-scale autonomous F1TENTH racing platform.

--- Page 4 ---

Figure 5: Data points obtained from the steady-state cornering experiment and the resulting model fit of the Pacejka model.

--- Page 7 ---

TABLE II: Vehicle Parameters Parameters Value Parameters Value

--- Page 3 ---

TABLE III: MAP2 Control Parameters Parameters Value Parameters Value

--- Page 5 ---

TABLE VI: Vehicle Parameters parameters m 4.24 kg [Bf, Br] [6.03, 18.67] lf 0.168 m [Df, Dr] [0.87, 0.87] lr 0.162 m [Ef, Er] [0.19, 0.27]

--- Page 5 ---

TABLE VI: Vehicle Parameters parameters m 4.24 kg [Bf, Br] [6.03, 18.67] lf 0.168 m [Df, Dr] [0.87, 0.87] lr 0.162 m [Ef, Er] [0.19, 0.27]<ref:2610.

Improvements for AI systems

  1. textbf Kinematic MPC for Predictive Lateral-Acceleration Generation: Replace geometric look-ahead with an MPC formulation that optimizes virtual inputs over a prediction horizon to enable predictive lateral acceleration planning for operation near the vehicle’s handling limits.

  2. textbf SGP-Based Compensation for Model Mismatch: Implement an SGP residual model to compensate for modeling errors by correcting the nominal steering command, as described by learning a steering residual from realworld driving data collected on the F1TENTH vehicle and calculating the corrected steering command before limiting is computed as δprek = δnom,k + ∆cδk.

  3. textbf Real-Time Residual Correction: Achieve computational efficiency for model correction by using a sparse formulation of SGP inference with a cost reduced to 2.0 ms, enabling the system to perform this correction in real-time on embedded racing platforms.

  4. textbf Robustness to Parameter Mismatch: Enhance controller robustness during cross-tire validation by using SGP's ability to adapt, as demonstrated by SGP further retrains the SGP on driving data logged on tire B when the offline LUT was generated using tire A's parameters only.

  5. textbf Enhanced High-Speed Tracking Performance: Improve tracking accuracy at high speeds by leveraging the combined approach, which shows that SGP correction reduces the average lateral acceleration tracking RMSE from 1.252 to 1.141 m/s2 and further improves performance compared to MAP2 without SGP.

Sources

Related papers