The reviewed record of science sign in
Pith

arxiv: 2605.08489 · v1 · submitted 2026-05-08 · cs.RO

LE-PAVD: Learning-Enhanced Physics-Aware Vehicle Dynamics for High-Speed Autonomous Navigation

Musabbir Ahmed Arrafi , Malik Ali , Nicholas M. Stiffler , Krishna Bhavithavya Kidambi This is my paper

Reviewed by Pith2026-05-12 02:05 UTCgrok-4.3open to challenge →

classification cs.RO
keywords vehicle dynamicsphysics-informed learningautonomous racinghybrid modelingtire forcesstate predictionclosed-loop controlmodel generalization
0
0 comments X

The pith

A hybrid model blends physics tire forces and load transfer with neural learning to predict vehicle states more accurately than deep networks alone.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that adding four specific physics components to an end-to-end trained neural dynamics model creates a more consistent and generalizable representation of vehicle behavior at high speeds. This matters because accurate dynamics are required for controllers operating near handling limits in autonomous racing, where purely learned models often lack reliability on new tracks and model-based methods oversimplify nonlinear effects. If the integration works as described, controllers could achieve lower prediction errors, faster lap times, and reduced computation without losing physical grounding. The authors train and evaluate the resulting model on both simulation and real telemetry data.

Core claim

LE-PAVD adds load-sensitive Pacejka tire forces, longitudinal load transfer, lateral tire-force effects, and rate-limited actuator inputs to a neural vehicle dynamics model. Trained end-to-end, the hybrid structure enforces physical consistency while learning nonlinearities from data, producing lower state prediction errors and better closed-loop performance than a deep dynamics baseline on both seen and unseen tracks.

What carries the argument

The integration of four physics components—load-sensitive Pacejka tire forces, longitudinal load transfer, lateral tire-force effects, and rate-limited actuator inputs—into an end-to-end neural dynamics model.

If this is right

  • On unseen tracks the model reduces average displacement error by 16.1 percent and final displacement error by 20.6 percent.
  • Yaw-rate root mean squared error drops 91.3 percent relative to a deep dynamics baseline.
  • The model requires 21.6 percent fewer FLOPs and runs about 1.5 times faster during inference.
  • Closed-loop simulations show 17.4 percent faster laps on a training track and 9.5 percent faster laps on a test track with no boundary violations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same pattern of adding targeted physics terms to neural models could be tested for other high-speed systems such as fixed-wing aircraft or underwater vehicles.
  • The efficiency gains suggest the approach may scale to onboard hardware with limited compute in production autonomous vehicles.
  • Further work could examine whether the same components improve robustness when the vehicle parameters change slightly, such as tire wear or payload shifts.

Load-bearing premise

That the four physics components can be added to a learned model in a way that keeps physical consistency, improves accuracy on new tracks, and avoids new inconsistencies or overfitting.

What would settle it

If the hybrid model produces higher average or final displacement errors, higher yaw-rate RMSE, or causes track boundary violations in closed-loop tests on an unseen track compared to the pure learned baseline, the claimed improvements would not hold.

Figures

Figures reproduced from arXiv: 2605.08489 by Krishna Bhavithavya Kidambi, Malik Ali, Musabbir Ahmed Arrafi, Nicholas M. Stiffler.

Figure 1
Figure 1. Figure 1: Single-track vehicle model used in LE-PAVD. State information shown as [𝑣𝑥 , 𝑣𝑦 , 𝜔] ⊤, along with slip angles 𝛼𝑓 and 𝛼𝑟 , normal loads 𝐹𝑓 𝑧 and 𝐹𝑟 𝑧 computed with longitudinal load transfer, and tire forces 𝐹𝑟 𝑥, 𝐹𝑓 𝑦, 𝐹𝑟 𝑦. This baseline structure is extended in our proposed framework. robustness under limited data. However, it adopts a sequential training approach rather than end-to-end optimization and… view at source ↗
Figure 2
Figure 2. Figure 2: LE-PAVD architecture. Sequential feedback and commanded control inputs are processed by a GRU and dense layers to predict next-step states [𝑣𝑥 , 𝑣𝑦 , 𝜔] ⊤. A Physics Guard Layer enforces parameter bounds and computes interpretable dynamics through slip angles, normal loads, and tire/longitudinal forces, and accelerations while capturing rate-limited actuator behavior. GRU network, followed by hidden dense … view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of predicted and ground truth [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of predicted and ground truth [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of predicted and ground truth [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison for 𝑣𝑥, 𝑣𝑦, and 𝜔 on ETHZMobil track MPC run. Deep Dynamics Model (DDM) in red, INA in brown and LE-PAVD model in green [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Internally estimated tire force components during closed-loop NMPC [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
read the original abstract

Accurate modeling of nonlinear vehicle dynamics is essential for high-speed autonomous racing, where controllers operate at the handling limits. Model-based methods are interpretable but rely on simplifying assumptions, while purely learned models capture nonlinearities yet often lack physical consistency and generalization. We propose LE-PAVD (Learning-Enhanced Physics-Aware Vehicle Dynamics), a hybrid model that integrates physics priors with learned components. Our architecture adds four components: load-sensitive Pacejka tire forces, longitudinal load transfer, lateral tire-force effects, and rate-limited actuator inputs. Trained end-to-end on simulation and real-world telemetry, LE-PAVD enforces physical consistency while improving state prediction accuracy. On an unseen track, LE-PAVD reduces average displacement error (ADE) by 16.1$\%$, final displacement error (FDE) by 20.6$\%$, and lowers yaw-rate root mean squared error (RMSE) by 91.3$\%$ versus a deep dynamics baseline, while using 21.6$\%$ fewer FLOPs and achieving approximately 1.50$\times$ faster inference. In closed-loop simulations, LE-PAVD consistently outperforms the baseline by achieving faster lap times by 17.4$\%$ on a training track and 9.5$\%$ on a test track, without any track boundary violations. Overall, LE-PAVD offers a compact, physics-grounded dynamics backbone that improves predictive fidelity and closed-loop performance while reducing inference cost.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes LE-PAVD, a hybrid vehicle dynamics model for high-speed autonomous racing that augments a learned backbone with four explicit physics components (load-sensitive Pacejka tire forces, longitudinal load transfer, lateral tire-force effects, and rate-limited actuator inputs). Trained end-to-end on simulation and real telemetry, the model claims to enforce physical consistency while delivering improved state prediction on unseen tracks (ADE reduced 16.1%, FDE 20.6%, yaw-rate RMSE 91.3%) and faster closed-loop lap times (17.4% training track, 9.5% test track) with lower FLOPs and faster inference versus a deep dynamics baseline.

Significance. If the reported gains and consistency enforcement hold, the work demonstrates a practical route to embedding targeted physics priors into end-to-end learned dynamics, yielding both better generalization to unseen tracks and reduced inference cost. This hybrid style could strengthen model-based controllers in handling-limit regimes where pure data-driven models often fail.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (Architecture): the central claim that the four physics components can be integrated while preserving physical consistency and avoiding new inconsistencies is load-bearing, yet the manuscript provides no explicit loss terms, constraint formulation, or residual-connection equations showing how consistency is enforced during end-to-end training.
  2. [§4] §4 (Experiments): the 16.1% ADE, 20.6% FDE, and 91.3% yaw-rate RMSE reductions on an unseen track are presented without error bars, number of trials, or data-split details; this weakens the generalization claim and makes it impossible to judge whether the gains are robust or sensitive to post-hoc tuning.
minor comments (2)
  1. Clarify the exact measurement protocol for the 'approximately 1.50× faster inference' and 21.6% fewer FLOPs claims, including hardware and batch size.
  2. Define ADE, FDE, and RMSE at first use in the abstract and main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed review. We address each major comment below and indicate the changes incorporated in the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (Architecture): the central claim that the four physics components can be integrated while preserving physical consistency and avoiding new inconsistencies is load-bearing, yet the manuscript provides no explicit loss terms, constraint formulation, or residual-connection equations showing how consistency is enforced during end-to-end training.

    Authors: We agree that an explicit formulation strengthens the presentation. The four physics components are integrated as fixed additive terms in the state-transition equations, with the learned backbone predicting residuals around these physics-based predictions. This architectural design enforces the priors in every forward pass without additional loss terms. In the revised manuscript we have added the residual-connection equations and a step-by-step description of the integration in Section 3 to clarify how consistency is maintained. revision: yes

  2. Referee: [§4] §4 (Experiments): the 16.1% ADE, 20.6% FDE, and 91.3% yaw-rate RMSE reductions on an unseen track are presented without error bars, number of trials, or data-split details; this weakens the generalization claim and makes it impossible to judge whether the gains are robust or sensitive to post-hoc tuning.

    Authors: The referee correctly notes the missing statistical details. In the revised Section 4 we now report error bars as standard deviation over five independent runs with distinct random seeds, specify the data splits (70 % training, 15 % validation, 15 % test from combined simulation and real telemetry), and confirm that the unseen-track evaluation uses a completely held-out track. These additions show the gains are consistent across trials. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper explicitly adds four physics components (load-sensitive Pacejka tire forces, longitudinal load transfer, lateral tire-force effects, and rate-limited actuator inputs) as priors into a hybrid architecture, then trains the full model end-to-end on simulation and real telemetry data. Evaluation metrics (ADE, FDE, yaw-rate RMSE, lap times) are reported on unseen tracks against a deep dynamics baseline, with no equations or claims showing that any prediction reduces by construction to the fitted inputs or to a self-citation. The central claims rest on empirical generalization results rather than tautological re-derivations, making the chain self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides insufficient detail to enumerate specific free parameters, axioms, or invented entities; the model relies on standard vehicle dynamics concepts such as the Pacejka tire model but does not introduce new postulated entities.

pith-pipeline@v0.9.0 · 5581 in / 1272 out tokens · 63751 ms · 2026-05-12T02:05:29.545994+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. CAPE: Control Algorithm Performance Evaluation under Learned Vehicle Dynamics Models

    eess.SY 2026-06 unverdicted novelty 5.0

    CAPE benchmark shows the proposed EPM learned dynamics model yields faster laps and lower errors than DPM and DDM across five controllers in both nominal and disturbance-aware simulations.

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · cited by 1 Pith paper

  1. [1]

    Autonomous robot racing competitions: Truly multivehicle autonomous racing competitions,

    H. Moon, S. H. Kang, J. Eom, M. J. Hwang, Y. Kim, J. Wang, B. Kim, T. Kim, T. Ga, J. Choi,et al., “Autonomous robot racing competitions: Truly multivehicle autonomous racing competitions,”IEEE Robotics & Automation Magazine, vol. 31, no. 1, pp. 123–132, 2024

    work page 2024

  2. [2]

    Hierarchical control for head-to-head autonomous racing,

    R. S. Thakkar, A. S. Samyal, D. Fridovich-Keil, Z. Xu, and U. Topcu, “Hierarchical control for head-to-head autonomous racing,”Field Robotics, vol. 4, pp. 46–69, 2024

    work page 2024

  3. [3]

    Autonomous driving small-scale cars: A survey of recent development,

    D. Li, P. Auerbach, and O. Okhrin, “Autonomous driving small-scale cars: A survey of recent development,”IEEE Transactions on Intelligent Transportation Systems, 2025

    work page 2025

  4. [4]

    Literature review and fundamental approaches for vehicle and tire state estimation,

    K. B. Singh, M. A. Arat, and S. Taheri, “Literature review and fundamental approaches for vehicle and tire state estimation,”Vehicle system dynamics, vol. 57, pp. 1643–1665, 2019

    work page 2019

  5. [5]

    H. B. Pacejka,Tire and Vehicle Dynamics. Amsterdam, The Netherlands: Elsevier, 2012

    work page 2012

  6. [6]

    Neural network vehicle models for high- performance automated driving,

    N. A. Spielberg et al., “Neural network vehicle models for high- performance automated driving,”Science Robotics, vol. 4, no. 28, p. eaaw1975, 2019

    work page 2019

  7. [7]

    Model- and acceleration-based pursuit controller for high-performance autonomous racing,

    J. Becker, N. Imholz, L. Schwarzenbach, E. Ghignone, N. Baumann, and M. Magno, “Model- and acceleration-based pursuit controller for high-performance autonomous racing,” inProc. IEEE Int. Conf. Robotics and Automation (ICRA), pp. 5276–5283, 2023

    work page 2023

  8. [8]

    Physics embedded neural network vehicle model and applications in risk-aware autonomous driving using latent features,

    T. Kim, H. Lee, and W. Lee, “Physics embedded neural network vehicle model and applications in risk-aware autonomous driving using latent features,” inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst., pp. 4182–4189, 2022

    work page 2022

  9. [9]

    Deep dynamics: Vehicle dynamics modeling with a physics-constrained neural network for autonomous racing,

    J. Chrosniak, J. Ning, and M. Behl, “Deep dynamics: Vehicle dynamics modeling with a physics-constrained neural network for autonomous racing,”IEEE Robotics and Automation Letters, vol. 9, pp. 5292 – 5297, 2024

    work page 2024

  10. [10]

    Fine-tuning hybrid dynamics with physics-informed neural networks for vehicle dynamics estimation,

    S. Fang and K. Yu, “Fine-tuning hybrid dynamics with physics-informed neural networks for vehicle dynamics estimation,”International Journal of Intelligent Robotics and Applications, pp. 1–17, 2025

    work page 2025

  11. [11]

    Guiggiani,The science of vehicle dynamics: handling, braking, and ride of road and race cars

    M. Guiggiani,The science of vehicle dynamics: handling, braking, and ride of road and race cars. Springer, 2018

    work page 2018

  12. [12]

    Tube model predictive control for an autonomous race car,

    A. Wischnewski, M. Euler, S. Gümüs, and B. Lohmann, “Tube model predictive control for an autonomous race car,”Vehicle System Dynamics, vol. 60, no. 9, pp. 3151–3173, 2022

    work page 2022

  13. [13]

    A survey of vehicle dynamics modeling methods for autonomous racing: Theoretical models, physical/virtual platforms, and perspectives,

    T. Zhang, Y. Sun, Y. Wang, B. Li, Y. Tian, and F.-Y. Wang, “A survey of vehicle dynamics modeling methods for autonomous racing: Theoretical models, physical/virtual platforms, and perspectives,”IEEE Transactions on Intelligent Vehicles, vol. 9, no. 3, pp. 4312–4334, 2024. 9

    work page 2024

  14. [14]

    An efficient minimum- time trajectory generation strategy for two-track car vehicles,

    A. Rucco, G. Notarstefano, and J. Hauser, “An efficient minimum- time trajectory generation strategy for two-track car vehicles,”IEEE Transactions on Control Systems Technology, vol. 23, no. 4, pp. 1505– 1519, 2015

    work page 2015

  15. [15]

    Deep neuralnetworkstopredictautonomousgroundvehiclebehavioronsloping terrain field,

    C. Badgujar, S. Das, D. M. Figueroa, D. Flippo, and S. Welch, “Deep neuralnetworkstopredictautonomousgroundvehiclebehavioronsloping terrain field,”Journal of Field Robotics, vol. 40, no. 4, pp. 919–933, 2023

    work page 2023

  16. [16]

    Learning dynamics models for velocity estimation in autonomous racing,

    J. Węgrzynowski, G. Czechmanowski, P. Kicki, and K. Walas, “Learning dynamics models for velocity estimation in autonomous racing,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 972–979, IEEE, 2024

    work page 2024

  17. [17]

    Deepracing: A framework for autonomous racing,

    T. Weiss and M. Behl, “Deepracing: A framework for autonomous racing,” in Proc. IEEE Des. Automat. Test Eur. Conf. Exhib., pp. 1163–1168, 2020

    work page 2020

  18. [18]

    Learning-based near-optimal motion planning for intelligent vehicles with uncertain dynamics,

    Y. Lu, X. Zhang, X. Xu, and W. Yao, “Learning-based near-optimal motion planning for intelligent vehicles with uncertain dynamics,”IEEE Robotics and Automation Letters, vol. 9, no. 2, pp. 1532–1539, 2023

    work page 2023

  19. [19]

    Digital twin of a driver- in-the-loop race car simulation with contextual reinforcement learning,

    S. Ju, P. van Vliet, O. Arenz, and J. Peters, “Digital twin of a driver- in-the-loop race car simulation with contextual reinforcement learning,” IEEE Robotics and Automation Letters, vol. 8, no. 7, pp. 4107–4114, 2023

    work page 2023

  20. [20]

    Learning based mpc for autonomous driving using a low dimensional residual model,

    Y. Li, C. Huang, D. Yang, W. Liu, and J. Li, “Learning based mpc for autonomous driving using a low dimensional residual model,”arXiv preprint arXiv:2412.03874, 2024

    work page arXiv 2024

  21. [21]

    A physics- informedapproachforlearningvehicledynamicsinhigh-speedautonomy,

    M. P. Kolluri, S. V. Shelar, S. A. Sarwar, and Z. Hasnain, “A physics- informedapproachforlearningvehicledynamicsinhigh-speedautonomy,” in AIAA SCITECH 2025 Forum, p. 1726, 2025

    work page 2025

  22. [22]

    Physics-informed adaptive deep koopman operator modeling for autonomous vehicle dynamics,

    J. Zhang, Y. He, and H. Chen, “Physics-informed adaptive deep koopman operator modeling for autonomous vehicle dynamics,”arXiv preprint arXiv:2503.23396, 2025

    work page arXiv 2025

  23. [23]

    Racecar-the dataset for high-speed autonomous racing,

    A. Kulkarni, J. Chrosniak, E. Ducote, F. Sauerbeck, A. Saba, U. Chirimar, J. Link, M. Behl, and M. Cellina, “Racecar-the dataset for high-speed autonomous racing,” in2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pp. 11458–11463, IEEE, 2023

    work page 2023

  24. [24]

    Bayesrace: Learning to race autonomously using prior experience,

    A. Jain, M. O’Kelly, P. Chaudhari, and M. Morari, “Bayesrace: Learning to race autonomously using prior experience,” inProc. 4th Conf. Robot Learn., pp. 1918–1929, 2020

    work page 1918

  25. [25]

    Social lstm: Human trajectory prediction in crowded spaces,

    A. Alahi, K. Goel, V. Ramanathan, A. Robicquet, L. Fei-Fei, and S. Savarese, “Social lstm: Human trajectory prediction in crowded spaces,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., pp. 961–971, 2016

    work page 2016