Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning dynamics models for velocity estimation in autonomous racing

Jan Węgrzynowski, Grzegorz Czechmanowski, Piotr Kicki, Krzysztof Walas

TL;DR

This work tackles velocity estimation under aggressive racing conditions by integrating a differentiable UKF with a hybrid dynamics model and online tire-road friction estimation. By optimizing dynamics-model parameters through state-estimation loss rather than pure one-step prediction loss, the approach improves estimation accuracy and enables zero-shot adaptation to new surfaces via online friction scaling. The method is validated on a publicly released F1TENTH dataset featuring extreme sideslip and high slip, and outperforms several learning-based state estimators while offering robust uncertainty handling through structured noise models. The combination of friction-aware dynamics, differentiable filtering, and a diverse dataset has practical implications for robust, real-time state estimation in high-speed autonomous racing.

Abstract

Velocity estimation is of great importance in autonomous racing. Still, existing solutions are characterized by limited accuracy, especially in the case of aggressive driving or poor generalization to unseen road conditions. To address these issues, we propose to utilize Unscented Kalman Filter (UKF) with a learned dynamics model that is optimized directly for the state estimation task. Moreover, we propose to aid this model with the online-estimated friction coefficient, which increases the estimation accuracy and enables zero-shot adaptation to the new road conditions. To evaluate the UKF-based velocity estimator with the proposed dynamics model, we introduced a publicly available dataset of aggressive manoeuvres performed by an F1TENTH car, with sideslip angles reaching 40°. Using this dataset, we show that learning the dynamics model through UKF leads to improved estimation performance and that the proposed solution outperforms state-of-the-art learning-based state estimators by 17% in the nominal scenario. Moreover, we present unseen zero-shot adaptation abilities of the proposed method to the new road surface thanks to the use of the proposed learning-based tire dynamics model with online friction estimation.

Learning dynamics models for velocity estimation in autonomous racing

TL;DR

This work tackles velocity estimation under aggressive racing conditions by integrating a differentiable UKF with a hybrid dynamics model and online tire-road friction estimation. By optimizing dynamics-model parameters through state-estimation loss rather than pure one-step prediction loss, the approach improves estimation accuracy and enables zero-shot adaptation to new surfaces via online friction scaling. The method is validated on a publicly released F1TENTH dataset featuring extreme sideslip and high slip, and outperforms several learning-based state estimators while offering robust uncertainty handling through structured noise models. The combination of friction-aware dynamics, differentiable filtering, and a diverse dataset has practical implications for robust, real-time state estimation in high-speed autonomous racing.

Abstract

Velocity estimation is of great importance in autonomous racing. Still, existing solutions are characterized by limited accuracy, especially in the case of aggressive driving or poor generalization to unseen road conditions. To address these issues, we propose to utilize Unscented Kalman Filter (UKF) with a learned dynamics model that is optimized directly for the state estimation task. Moreover, we propose to aid this model with the online-estimated friction coefficient, which increases the estimation accuracy and enables zero-shot adaptation to the new road conditions. To evaluate the UKF-based velocity estimator with the proposed dynamics model, we introduced a publicly available dataset of aggressive manoeuvres performed by an F1TENTH car, with sideslip angles reaching 40°. Using this dataset, we show that learning the dynamics model through UKF leads to improved estimation performance and that the proposed solution outperforms state-of-the-art learning-based state estimators by 17% in the nominal scenario. Moreover, we present unseen zero-shot adaptation abilities of the proposed method to the new road surface thanks to the use of the proposed learning-based tire dynamics model with online friction estimation.
Paper Structure (18 sections, 9 equations, 6 figures, 2 tables)

This paper contains 18 sections, 9 equations, 6 figures, 2 tables.

Table of Contents

  1. INTRODUCTION
  2. RELATED WORK
  3. MATERIALS AND METHODS
  4. Problem definition
  5. Proposed solution
  6. UKF update
  7. UKF prediction
  8. Noise model
  9. Model training procedure
  10. Friction coefficient estimation
  11. Dataset
  12. EXPERIMENTS
  13. Experimental setup
  14. Velocity estimation in known conditions
  15. Zero-shot generalization to new conditions
  16. ...and 3 more sections

Figures (6)

  • Figure 1: The main goal of this paper is to learn a dynamics model of an F1/10 car for accurate velocity estimation during aggressive maneuvers.
  • Figure 2: The general scheme of the differentiable -based state estimation, with learnable vehicle and noise models. TODO: xk+1 na wejsciu do lossa powinno byc
  • Figure 3: Dataset Distribution Comparison: Our Data vs. Maximum Values from End-to-End Velocity Estimation srinivasan2020endtoend, Robust LSTM ghosn2023lstm, and 99th Percentile of Targa Sixty-Six Dataset TargaSixtySix.
  • Figure 4: XXXXTODO: placeholder - create better plot, change titles
  • Figure 5: Comparison of estimation performance on unseen during training surface.
  • ...and 1 more figures