REVIEW 4 major objections 5 minor 20 references
ZeloS -- A Research Platform for Early-Stage Validation of Research Findings Related to Automated Driving
T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash
Pith's one-line read This paper presents ZeloS, a 69-kg modular research platform for early-stage validation of automated driving methods, and shows experiments in which it navigates to a target and avoids obstacles.
desk verdict A well-engineered modular research platform whose central validation claim outruns the experimental evidence: one qualitative run supports functioning, not reliability. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the constant coupling matrix $G$ (Eq. 2): because the steering axes are located exactly above the tire contact points and the suspension is linear, the contact points stay fixed relative to the center of gravity, so one constant $3 \times 8$ matrix maps tire forces to the COG force and torque vector, and its transpose maps COG velocity to the eight tire velocities. The Moore-Penrose inverse $G^+$ of $G$ carries both the encoder-based odometry (Eq. 12) and the analytical control allocation (Eq. 25), so the platform's model, localization, and control all rest on this single geometric idealization. Around that, the automation uses a convex model predictive planner that linearizes the dynamics along the previous solution and restricts the trajectory to a convex collision-free polytope built from LIDAR point clouds, plus a 100 Hz tracking MPC with feedforward velocity and acceleration and a PI velocity controller.
What would settle it
Measure the tire contact point positions relative to the COG while sweeping steering angles and deflecting the suspension, or run the vehicle through aggressive maneuvers with tire slip and compare the encoder-based odometry (Eq. 12) against an external motion capture ground truth. If the odometry error grows in a way the constant-$G$ model cannot explain, or if the force allocation produces unintended yaw moments, the central assumption is falsified.
Extended reading notes
Core claim
The paper's central claim is that a deliberately simple, modular vehicle design can serve as a versatile research platform for validating automated driving methods at an early stage. Each of the four drive modules carries its own actuator, power, and computer, so the wheel arrangement and steering mode can be reconfigured, and the modular automation lets individual modules such as sensors, perception, planning, and control be swapped. The implemented baseline uses a constant coupling matrix G derived from the chassis geometry to relate tire forces and velocities to the center of gravity, enabling a compact dynamic model. Motion planning is a model predictive planner that linearizes the dynamics and constrains the trajectory to a convex, collision-free polytope extracted from LIDAR point clouds, and tracking control is a 100 Hz MPC cascaded with a PI velocity controller and analytical control allocation. Experimental results in the paper show position tracking errors below 20 mm, planner and controller cycle times meeting their targets, and a scenario where the platform avoids an obstacle and navigates from start to goal.
Load-bearing premise
The load-bearing premise is that the tire contact points remain fixed relative to the center of gravity because the steering axes sit exactly above the contact points and the suspension is linear; if this geometric idealization fails, the constant coupling matrix and the odometry and force allocation built on it become biased.
Editorial extensions
If this is right
- New motion planning or control modules can be validated on ZeloS by replacing only the relevant automation module, because the interfaces are modular and self-contained.
- The convex optimization formulation with linearized dynamics runs in real time on the onboard computer, with average planner cycle time 37.54 ms and MPC cycle time 5.43 ms, and more than 99.7% of MPC cycles met the target cycle time.
- Position tracking error stayed below 20 mm in the presented scenario, indicating that the platform can support demanding control validations.
- The obstacle-avoidance experiment shows that the planner can keep the reference trajectory inside a convex, collision-free admissible set computed from LIDAR data.
- The digital twin provides software-in-the-loop testing with the same interfaces as the real vehicle, allowing fast and risk-free checks before hardware experiments.
Reading between the lines
- A natural extension is to quantify when the constant-coupling assumption degrades: under high tire slip or suspension deflection, odometry and force allocation will be biased, and comparing encoder odometry against an external motion capture system during aggressive maneuvers would reveal the operating envelope.
- The modular architecture suggests ZeloS could serve as a shared benchmark platform on which different research groups swap planning or control modules and compare methods on identical hardware.
- Given the paper's stated motivation of validating formally safe driving functions, the small scale and low risk make ZeloS a plausible testbed for safety-certified planning and control before full-scale vehicle tests.
- A testable extension is to use the digital twin to inject tire slip and suspension deflection, then check whether the tracking error and force allocation remain acceptable compared with the real vehicle, isolating the effect of the constant-coupling idealization.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper presents ZeloS, a 69 kg, 117 cm research platform for early-stage validation of automated driving methods. It describes the modular hardware (all-wheel steering, all-wheel drive, distributed computation) and a centralized ROS-based automation architecture, then focuses on the implemented motion planner and tracking controller: an MPC-based path-following planner that uses convex polytopic free space from the 2D LIDAR, a cascaded MPC/PI tracking controller, and an analytical force-allocation scheme. The experimental section reports one delivery scenario, for which the tracking error is stated to be below 20 mm, and one obstacle-avoidance scenario shown as a planning snapshot at t = 2 s. The authors conclude that ZeloS navigates reliably and autonomously avoids obstacles.
Significance. If the platform performs as claimed, this is a useful engineering contribution: it provides a moderately sized, modular, sensor-rich testbed for early validation of motion planning and control algorithms, with a digital twin for software-in-the-loop testing. The paper is clearly written, gives a fairly complete hardware and software architecture, and models the vehicle dynamics and control allocation in enough detail to be reproduced. The reported real-world run helps demonstrate that the full stack works end to end. However, the quantitative support for the central claim is currently thin: there is only a single experimental run and a single planning snapshot, with no statistical analysis, repeated trials, or avoidance-completion metrics. The central claim of a functional platform is defensible, but the evidence as presented is not yet sufficient to support the words 'reliably' and 'autonomously avoids obstacles'.
major comments (4)
- [Section VII and Section VIII] The central conclusion in Section VIII that ZeloS 'navigates reliably' and 'autonomously avoids obstacles' is not supported by the reported experiments. Section VII presents one delivery run, with tracking error summarized in Fig. 11 only as 'at most < 20 mm', and one obstacle-avoidance scenario shown as a planning snapshot at t = 2 s in Fig. 12. There are no repeated trials, no success/failure counts, no error statistics or confidence intervals, and no measure of obstacle clearance or maneuver completion. Reliability is a statistical property; a single run cannot establish it. The obstacle-avoidance figure shows only the planned reference trajectory at one instant, not whether the executed vehicle cleared the obstacle. The authors should either add repeated trials and completion/clearance metrics for both scenarios, or substantially weaken the conclusion to a demonstration of a single successful run and a single planning instance.
- [Section V-C1, Eq. (16)] Equation (16) as printed applies the rotation matrix only to the reference vector and then subtracts the measured position: xε = R(φhat) [xr yr φr]^T - [xhat yhat φhat]^T. With xr, xhat in the global frame, this expression is not a body-frame error and is not invariant to absolute position; it is also inconsistent with the standard error dynamics in Eq. (17), which are derived from e = R(-φhat)(p_r - p_hat). The correct form should be xε = R(-φhat)( [xr yr]^T - [xhat yhat]^T ) for the translational part. Please correct Eq. (16) or clarify the coordinate definitions that would make it valid; if the implementation uses the standard form, the paper should state so.
- [Section II-B, Eq. (2)] The constant coupling matrix G in Eq. (2) rests on the assumption that tire contact points remain fixed relative to the COG for arbitrary steering angles and suspension deflections, because the steering axes are located exactly above the contact points and the suspension is linear. This assumption is used in odometry (Eq. (12)), force allocation (Eq. (25)), and dynamics (Eq. (4)). The paper offers no direct experimental check of this assumption. Given that the single reported run achieved <20 mm tracking error, the assumption may be acceptable for that scenario, but a targeted validation (e.g., comparing encoder-based odometry against the Vicon ground truth under a steering-angle sweep and under varying suspension loads) would de-risk the model for the broader claim that ZeloS supports validation of various automated driving methods.
- [Section V-B and V-C1, Eqs. (15e) and (21a)] The optimization problems that define the planner and the position MPC are not reproducible from the paper because several tuning parameters are omitted: the MPC weight matrices Q, S, R in Eq. (21a), the reinitialization threshold vector xth_epsilon in Eq. (15e), and the polytopic sets V_o and A_o in Eqs. (21c)-(21d). These parameters directly affect the reported computation times and tracking performance. Please report their values or provide a repository or online appendix with the complete implementation.
minor comments (5)
- [Section IV-B, Eq. (11)] Equation (11) has a typesetting issue in the bracket structure; the formula should read \hat v_i = [\hat\omega_i r_dyn cos(\hat\delta_i), \hat\omega_i r_dyn sin(\hat\delta_i)]^T.
- [Section IV-B and Section V-C3, Eqs. (12) and (25)] The notation G+ is used both for the pseudoinverse of G in Eq. (25) and, in Eq. (12), for the pseudoinverse of G^T. Because these matrices have different dimensions, the notation should be disambiguated, e.g., (G^T)^+ in Eq. (12).
- [Figure 12 caption] The caption says 'ZeloS' position is at the green mark' while the green line denotes the reference trajectory; please disambiguate the marker and the line in both the caption and the legend.
- [Section II-C] The statement that the steering actuator reaches a desired tire angle 'almost instantly' is not quantified; a rise time or bandwidth would substantiate the modeling assumption in Section III that actuator dynamics can be neglected.
- [Section V-C2, Eq. (23)] The PI gains RP and RI are not specified. Since they are part of the implemented velocity control, reporting their values (or referencing accessible code) would improve reproducibility.
Circularity Check
No significant circularity; the platform demo is validated experimentally, with only a minor self-citation in the velocity control module.
full rationale
ZeloS is a systems-and-experiments paper rather than a derivation paper. The control stack imports published methods: the tire model and control allocation from [12] (Moseberg), the velocity control from [16] (Schwartz/Rudolf/Hohmann, with Hohmann as a co-author of the present paper), the bi-level planning architecture from [15], and the convex-polytope generation from [14]. None of these imports is presented as a new result or as a prediction derived inside this paper. The central claim—that ZeloS navigates reliably and avoids obstacles—is supported by the Section VII experimental data rather than by the cited equations themselves. In particular, Eqs. (22) and (25) transfer prior published control designs, but these designs are then implemented and exercised in a real delivery scenario with a reported tracking error below 20 mm and an obstacle-avoidance planning snapshot, so the experimental outcome is independent of the citation chain. The constant coupling matrix G in Eq. (2) follows from the explicitly stated mechanical design (steering axes located exactly above tire contact points and linear suspension), not from fitting to the demo. The odometry in Eq. (12) uses G+, but the localization is corrected by LIDAR-based SLAM, so the reported tracking error is not merely the controller being evaluated against its own fitted output. The only self-citation is [16], and it is a normal reuse of a prior external control design; it does not supply the evidence for the platform's performance. There are no fitted parameters renamed as predictions, no uniqueness theorem imported from the authors, and no ansatz smuggled in via citation. The experimental evidence is sparse (one run, no statistical repetition), and the conclusion is stronger than the data alone can support, but that is an evidence-strength and correctness concern, not circularity. Accordingly, the score reflects only a minor, non-circular self-citation.
Assumptions & free parameters
free parameters (3)
- MPC weight matrices Q, S, R (position control)
- Reinitialization threshold vector xth_epsilon
- Polytopic constraint sets V_o and A_o
assumptions (5)
- domain assumption Flat surface driving assumption
- domain assumption Neglect of actuator dynamics
- domain assumption Pure rolling for odometry
- domain assumption Constant coupling matrix G due to mechanical design
- domain assumption Simplified Magic Formula tire model
Cite this review
Pith. "Pith review of ZeloS -- A Research Platform for Early-Stage Validation of Research Findings Related to Automated Driving." pith.science (2026). https://pith.science/paper/4VSP6WL4
@misc{pith2026250502460,
author = {Pith},
title = {Pith review of: ZeloS -- A Research Platform for Early-Stage Validation of Research Findings Related to Automated Driving},
year = {2026},
howpublished = {\url{https://pith.science/paper/4VSP6WL4}},
note = {Machine review of arXiv:2505.02460}
}
read the original abstract
This paper presents ZeloS, a research platform designed and built for practical validation of automated driving methods in an early stage of research. We overview ZeloS' hardware setup and automation architecture and focus on motion planning and control. ZeloS weighs 69 kg, measures a length of 117 cm, and is equipped with all-wheel steering, all-wheel drive, and various onboard sensors for localization. The hardware setup and the automation architecture of ZeloS are designed and built with a focus on modularity and the goal of being simple yet effective. The modular design allows the modification of individual automation modules without the need for extensive onboarding into the automation architecture. As such, this design supports ZeloS in being a versatile research platform for validating various automated driving methods. The motion planning component and control of ZeloS feature optimization-based methods that allow for explicitly considering constraints. We demonstrate the hardware and automation setup by presenting experimental data.
Reference graph
Works this paper leans on
-
[16]
M. Schwartz, T. Rudolf, and S. Hohmann, “Robust position and velocity tracking control of a four-wheel drive and four-wheel steered electric vehicle,” in 2020 IEEE Int. Conf. Control, Automat. Robot. (ICCAR), 2020, pp. 415–422
work page 2020
-
[1]
Werling, Optimale aktive Fahreingriffe f¨ur Sicherheits- und Kom- fortsysteme in Fahrzeugen
M. Werling, Optimale aktive Fahreingriffe f¨ur Sicherheits- und Kom- fortsysteme in Fahrzeugen. Berlin: De Gruyter Oldenbourg, 2017
work page 2017
-
[2]
Porter, Accelerated Testing and Validation
A. Porter, Accelerated Testing and Validation. Amsterdam ; Boston, Mass: Newnes, 2004, 256 pp
work page 2004
-
[3]
ISO 26262: Road vehicles - Functional safety,
International Organization for Standardization, “ISO 26262: Road vehicles - Functional safety,” ISO, Geneva, Switzerland, Tech. Rep. ISO 26262:2018, 2018
work page 2018
-
[4]
Autonomous drifting using torque vectoring: Innovating active safety,
B. Lenzo, T. Goel, and J. Christian Gerdes, “Autonomous drifting using torque vectoring: Innovating active safety,” IEEE Trans. Intell. Transp. Syst., vol. 25, no. 11, pp. 17 931–17 939, 2024
work page 2024
-
[5]
P. Karle et al. , EDGAR: An autonomous driving research platform – from feature development to real-world application , 2024. arXiv: 2309.15492
arXiv 2024
-
[6]
Small-scale self-driving cars: A systematic literature review,
F. Caleffi, L. d. S. Rodrigues, J. d. S. Stamboroski, and B. M. Pereira, “Small-scale self-driving cars: A systematic literature review,” J. Traffic Transp. Eng., vol. 11, no. 2, pp. 271–292, 2024
work page 2024
-
[7]
B. D. Evans et al. , Unifying f1tenth autonomous racing: Survey, methods and benchmarks , 2024. arXiv: 2402.18558
arXiv 2024
Show all 20 references
-
[8]
AMZ Driverless: The full autonomous racing system,
J. Kabzan et al. , “AMZ Driverless: The full autonomous racing system,” J. Field Robot. , vol. 37, no. 7, pp. 1267–1294, 2020
2020
-
[9]
Towards a viable autonomous driving research plat- form,
J. Wei et al. , “Towards a viable autonomous driving research plat- form,” in 2013 IEEE Int. Vehicles Symp. (IV) , 2013, pp. 763–770
2013
-
[10]
Heinrich et al
M. Heinrich et al. , CoCar NextGen: A multi-purpose platform for connected autonomous driving research , 2024. arXiv: 2404 . 17550
2024
-
[11]
ROS: An open-source Robot Operating System,
Quigley, Morgan et al. , “ROS: An open-source Robot Operating System,” in IEEE Int. Conf. Robot. Automat. (ICRA) , 2009
2009
-
[12]
Moseberg, Regelung der Horizontalbewegung eines ¨uberak- tuierten Fahrzeugs unter Ber ¨ucksichtigung von Realisierungsan- forderungen
J.-E. Moseberg, Regelung der Horizontalbewegung eines ¨uberak- tuierten Fahrzeugs unter Ber ¨ucksichtigung von Realisierungsan- forderungen. FAU University Press, 2016
2016
-
[13]
Real-time loop closure in 2d lidar slam,
W. Hess, D. Kohler, H. Rapp, and D. Andor, “Real-time loop closure in 2d lidar slam,” in 2016 IEEE Int. Conf. Robot. Automat. (ICRA) , 2016, pp. 1271–1278
2016
-
[14]
Zhong et al
X. Zhong et al. , Generating Large Convex Polytopes Directly on Point Clouds, 2020. arXiv: 2010.08744
2020 arXiv
-
[15]
Werling, Ein Neues Konzept f ¨ur die Trajektoriengenerierung und -Stabilisierung in Zeitkritischen Verkehrsszenarien
M. Werling, Ein Neues Konzept f ¨ur die Trajektoriengenerierung und -Stabilisierung in Zeitkritischen Verkehrsszenarien . KIT Scientific Publishing, 2011
2011
-
[17]
On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming,
A. W ¨achter and L. T. Biegler, “On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear programming,” Math. Program., vol. 106, no. 1, pp. 25–57, 2006
2006
-
[18]
HPIPM: A high-performance quadratic programming framework for model predictive control *,
G. Frison and M. Diehl, “HPIPM: A high-performance quadratic programming framework for model predictive control *,” IFAC- PapersOnLine, 21st IFAC World Congress, vol. 53, no. 2, pp. 6563– 6569, 2020
2020
-
[19]
CasADi: A software framework for nonlinear optimization and optimal con- trol,
J. Andersson, J. Gillis, G. Horn, J. Rawlings, and M. Diehl, “CasADi: A software framework for nonlinear optimization and optimal con- trol,” Math. Program. Comp. , vol. 11, no. 1, pp. 1–36, 2019
2019
-
[20]
Design and use paradigms for gazebo, an open-source multi-robot simulator,
N. Koenig and A. Howard, “Design and use paradigms for gazebo, an open-source multi-robot simulator,” in IEEE/RSJ Int. Conf. Intell. Robots and Syst. , Sendai, Japan, 2004, pp. 2149–2154
2004
Reviewed August 16, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.