From perception to control: an autonomous driving system for a formula student driverless car

Huiqian Li; Tairan Chen; Yiting He; Zewen Xu; Zhe Yan; Zirui Li

REVIEW 3 major objections 6 minor 8 references

From perception to control: an autonomous driving system for a formula student driverless car

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that redundant cone perception plus model-predictive control lets a Formula Student racecar lap an unknown track faster and more precisely than pure pursuit.

desk verdict A useful FSAC system report whose headline MPC-vs-pure-pursuit claim is undercut by a non-matched baseline and single-run statistics. read the letter →

arxiv 1909.00119 v1 pith:IOCUCMP7 submitted 2019-08-31 cs.RO cs.CV

Tairan Chen , Zirui Li , Yiting He , Zewen Xu , Zhe Yan , Huiqian Li This is my paper · ORCID

classification cs.ROcs.CV

keywords autonomousdrivingFormulaStudentDriverlessconedetectionLiDARperceptionmonocularvisionExtendedKalmanFilteroccupancygridmapmodelpredictivecontrol

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper reports the autonomous-driving software stack of the "Smart Shark II" Formula Student racecar that won the 2018 FSAC competition. The aim is to show that a student-built racecar can complete two laps of a previously unknown, cone-bounded track at competitive speed and with safety, using redundant perception and predictive control. The authors claim that fusing LiDAR-based and vision-based cone detection, EKF localization, and an occupancy grid map gives stable perception, and that an MPC path-following controller keeps a higher average speed and lower lateral error than the pure-pursuit controller used in the competition. A sympathetic reader would care because the paper offers a working, integrated example of an autonomous racecar architecture built from off-the-shelf algorithms.

What carries the argument

The central object is a model predictive controller built on a 3-DOF vehicle model whose state includes position, heading, longitudinal and lateral speed, yaw rate, steering angle, acceleration, and lateral and heading errors. It predicts future states over a horizon while respecting constraints on steering angle, longitudinal speed, acceleration, steering rate, jerk, rear-tire sideslip angle, and a lateral feasible region derived from the track map, then solves the resulting nonlinear program with an hp-adaptive pseudospectral method. A supporting mechanism is the occupancy grid map, which accumulates cone detections from both LiDAR and camera over time and space so that single-frame perception errors do not destabilize the planned path.

What would settle it

Place both controllers on the same 4 m-wide circular track for many repeated laps, recording lateral error and average speed each lap; if the predictive controller's advantage shrinks or reverses across laps, the single-run comparison fails. Then, on a banked or uneven surface, compare the vision-based cone positions obtained from the flat-ground perspective transform with LiDAR-derived positions; disagreement beyond the 0.1 m grid resolution would falsify the flat-ground pose assumption.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the proposed MPC controller outperforms pure pursuit on the same 4 m-wide circular track: it maintains an average speed of $2.9720\ \mathrm{m/s}$ versus $1.3677\ \mathrm{m/s}$ for pure pursuit, while reducing average lateral path-following error from $0.4520\ \mathrm{m}$ to $0.2714\ \mathrm{m}$, standard deviation of lateral acceleration from $0.2338\ \mathrm{m/s^2}$ to $0.1759\ \mathrm{m/s^2}$, and average sideslip angle from $0.0120\ \mathrm{rad}$ to $0.0018\ \mathrm{rad}$. The controller achieves this with a 3-DOF vehicle model, steering-rate and jerk inputs, and constraints on steering, speed, acceleration, rear-tire sideslip, and a lateral corridor. The paper also reports that the EKF keeps the car localized after deliberate LiDAR failure, and that the occupancy grid map compensates for frame-level red/blue cone misclassification.

Load-bearing premise

The strongest quantitative claim sits on a single comparison run on a flat circular track with no repeated trials and no error bars; if that one run or the flat-ground pose assumption is unrepresentative of real circuits, the reported advantages would not stand.

Editorial extensions

If this is right

An FSAC-style racecar using the proposed controller can score higher in competitions, since higher average speed is the direct scoring quantity and the reported MPC run already laps a 4 m-wide circular track at more than double the pure-pursuit speed.
The localization module survives the loss of any single source because the EKF fuses LiDAR odometry, GPS/INS, and wheel-speed information, and the authors demonstrate continued stable localization after deliberate LiDAR failure.
Accumulating cone observations in a 0.1 m occupancy grid over time and space corrects the CNN's tendency to swap red and blue cone labels at turns, making single-frame misclassifications non-fatal.
The explicit rear-tire sideslip constraint keeps the vehicle in the linear tire regime, which explains the reported near-zero average sideslip angle while cornering at speed.
The modular architecture, with separate perception, mapping, localization, and control layers, can serve as a reference design for other Formula Student Driverless teams.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the reported single-track comparison is representative, an immediate test is to race both controllers on several different track shapes, such as hairpins, chicanes, and banked corners, recording per-lap error and speed; the MPC advantage should persist across all shapes if it comes from prediction rather than track geometry.
The flat-ground assumption in the vision pose estimation could be removed by estimating a local ground plane from the LiDAR point cloud before applying the perspective transform, which would extend the vision module to banked or uneven race tracks.
The occupancy-grid update currently uses only binary occupied/blank states; feeding the color confidence output of YOLOv3 into the log-odds update would give a soft, probabilistically principled way to fuse color evidence across frames.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Quick take: this is a reasonable system paper from the team that won FSAC 2018, but the headline quantitative claim about MPC beating pure pursuit should not be cited as evidence without a much more careful comparison. The reader's conditional verdict and the stress-test note both land; I agree with them.

What's actually new: the LiDAR-only cone color estimation, which turns a color classification problem into a geometric distribution problem and uses a small CNN over sorted cone positions. That's a neat trick, and it's trained in simulation and tested on real data, which is more than many competition papers do. The overall architecture—redundant LiDAR/vision perception, EKF localization, occupancy grid accumulation, MPC for longitudinal+lateral control—is a clean integration and is described at a level that would help another FSD team.

Where it's soft, in proportion: the MPC-vs-pure-pursuit experiment in Section III-C is the load-bearing evidence and it is confounded. Pure pursuit only steers; its speed comes from a separate planner. The paper doesn't report the pure pursuit speed reference, lookahead distance, or any controller gains. Pure pursuit's average speed is 1.37 m/s versus MPC's 2.97 m/s, yet it has higher lateral error and higher sideslip. That pattern suggests the baseline was not tuned or was run at a very conservative speed; a well-tuned pure pursuit at lower speed should typically track more accurately. Without repeated trials, error bars, or a matched speed profile, 'MPC can keep a higher speed with lower error' is underdetermined. Also, the MPC formulation has malformed equations (the A and B matrices in Eq. 18 look wrong, and the cost function in Eq. 28 has a typo), and no weights, horizon, or constraint values are given. The CNN red-blue confusion of 23.2% at 10 m is high; the paper says the occupancy grid compensates but doesn't show it quantitatively.

Bottom line: as a system description, it's useful and the authors are honest about what is simulation-trained and what is real. As a validation of MPC superiority, it doesn't hold up yet. I'd send it to peer review for the system and the CNN idea, but require a proper experimental comparison before the MPC claim is accepted. Cite it if you work on FSD system integration; don't cite it for the controller comparison.

Referee Report

3 major / 6 minor

Summary. The paper describes the autonomous driving software system of the 'Smart Shark II' Formula Student driverless car that won FSAC 2018. The architecture comprises EKF-based localization fusing LiDAR odometry, GPS/INS, and wheel speed; LiDAR-based cone detection with a CNN for color/geometry classification; vision-based cone detection with YOLOv3 and joint LiDAR-camera calibration; an occupancy grid map; and a model predictive controller for combined longitudinal and lateral path following. Real-world experiments on a circular track are presented, including a sensor-failure localization test, a cone classification comparison, and an MPC-vs-pure-pursuit comparison.

Significance. If the reported results are supported, the paper is a useful systems integration reference for Formula Student Driverless and similar low-speed autonomous racing platforms. It explicitly documents a redundant perception scheme, a practical mapping approach, and an MPC formulation that could inform other competition teams. The explicit architecture and real-world experimental data are strengths, and the EKF sensor-failure experiment is a concrete robustness demonstration. The central quantitative claim of MPC superiority, however, rests on a single-run comparison with incomplete experimental detail, so the significance of that specific claim is currently limited.

major comments (3)

[III-C, Table II] The headline comparison is not yet supported. Table II reports one run per controller, with no repeated trials, no error bars, and no statistical test. The comparison is also confounded: the MPC output includes longitudinal speed planning, whereas pure pursuit only sets the steering angle and the speed is commanded by a separate reference. The reported average speeds differ by more than a factor of two (2.9720 vs 1.3677 m/s), so the lower lateral error and sideslip of MPC could reflect the speed reference used for the pure pursuit baseline rather than a property of the path-following controller. To support the claim in Section III-C, the authors should report the pure pursuit lookahead and speed reference, the MPC horizon/weights/constraints, and results from repeated trials (or at minimum error bars from multiple laps) with matched speed planning.
[II-E, Eqs. (18)-(28)] The MPC formulation is underspecified and contains printed inconsistencies. The state vector ξ in Eq. (18) has ten components, but the displayed f, A, and B are not dimensionally consistent (e.g., A appears to have only two rows, and u1 is given as a three-element vector with a misplaced transpose). Equation (28) has a malformed third term, '∑ W_sh S_sh^(k) W_sh', and S_sh is never defined. Moreover, none of the weights W_u, W_eψ, W_ey, W_sh or the bounds in Eqs. (21)-(26) are given, and the prediction horizon and pseudospectral discretization parameters are omitted. Without these details the experiment cannot be reproduced, and the comparison in Table II cannot be attributed specifically to the MPC design.
[III-B, Fig. 15] The LiDAR CNN color classification is reported as 76.7573% correct within 10 m, and the text states that the CNN 'will reverse the classification of red and blue cones at 23.2% in the range of 10m'. This is a large systematic red/blue confusion rate that could flip the assignment of left/right track boundaries. The claim that 'this shortcoming can be overcome by constructing an occupancy grid map and accumulating the perceived results in time and space' is not quantified anywhere in the paper. Please provide a confusion matrix and a before/after comparison of map-level classification accuracy, and state the effect on the final trajectory.

minor comments (6)

[II-A, Eqs. (5)-(9)] The Kalman gain in Eq. (7) uses P_k while the prediction covariance is defined as P_k^hat in Eq. (6), and the text states that 'f is a three-dimensional dynamic model' although Eq. (2) defines a six-dimensional state; please standardize notation.
[II-B] The CNN input is described as a 30-dimensional vector, but the ordering of the 15 cone coordinates and the handling of fewer than 15 cones are not specified; please clarify the fixed ordering and padding rule.
[II-E] The variable κ in u2=[0, κ] is undefined, and the control input u1 appears with a misplaced transpose; please define all symbols in Table I.
[III-B, Fig. 15] The legend of Fig. 15 includes a 'Rule-Based' curve that is not described in the text; please explain this method and the source of ground-truth labels for the real-world dataset.
[III-C] The discussion of 'sample 140' attributes a large pure-pursuit error to 'poor controller response and inaccurate positioning' without supporting evidence; this speculation should be removed or replaced with a concrete analysis.
[Throughout] Minor typographical issues (e.g., 'which ensure high vehicle speed and safety', 'V ehicle') should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the MPC comparison is an experimental claim, and any baseline-tuning concerns are validity issues, not circular derivation.

full rationale

Walking the paper's claimed derivation chain: the LiDAR cone-color CNN is trained on simulated cone layouts and tested on real-world datasets not used for training; the vision YOLOv3 detector is trained on labeled real images and evaluated on held-out data. These are external benchmarks, not predictions forced by construction. The EKF localization is a standard sensor-fusion formulation with a stated nonlinear dynamic model and measurement model; the LiDAR-failure experiment demonstrates redundancy rather than deriving the fused output from the failure event. The occupancy grid map uses the standard log-odds Bayesian update with no parameters retrofitted to the reported map output. The MPC path-following controller minimizes a cost over steering rate, heading deviation, lateral error, and slack variables; the comparison against pure pursuit is an independent baseline, and the reported lower lateral error, lower sideslip, and higher average speed are experimental measurements. The pure-pursuit comparison may be confounded by separate speed planning, unreported controller tuning, and lack of repeated trials, but that is an experimental-validity concern, not circularity: no equation or citation in the paper reduces the claimed result to a fitted parameter, a renamed input, or a self-citation. No load-bearing self-citations or imported uniqueness theorems were found. The paper is therefore not circular; score 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claims depend on unstated MPC tuning parameters, the flat-ground assumption, the FSAC geometric color rule, and a linear tire model. No new physical entities are introduced.

free parameters (5)

MPC cost weights W_u, W_ey, W_eψ, W_sh
Used in the cost function (Eq. 28); no values are reported, and the comparison results depend on them.
Constraint bounds δ_f,min/max, U_min/max, a_min/max, ς_f,min/max, J_x,min/max
Defined in Eqs. (21)-(25) but numerical values are not given.
Sideslip angle limit α_r,lim
Used in constraint (26); value not reported.
Comfort distance d_s
Used in environmental constraint (27); value not reported.
CNN input padding size (15 cones) = 15
The CNN input is fixed to 15 cone coordinates with zero-padding; the actual number of cones in the FSAC track may vary and affects the geometric distribution.

assumptions (4)

domain assumption Ground is flat for vision-based pose estimation (Section II-C-2).
The homography between the image plane and the ground plane assumes a planar ground surface.
domain assumption FSAC cone placement rule: left and right track boundaries have different colors, making color estimable from geometric distribution (Section II-B-2).
The CNN's input is only cone coordinates; the classification relies on the rule that red cones are on one side and blue on the other.
domain assumption Linear tire model with rear-slip constraint keeps tires in the linear range (Section II-E-2).
The MPC uses a linear tire model and a constraint to avoid saturation, which may not hold at high speed or on slippery surfaces.
standard math The EKF motion model (Eq. (3)) assumes constant acceleration within a sampling step and does not model sensor biases.
These are typical EKF assumptions; the paper does not provide evidence that the racecar dynamics fit this model.

how reviews work

0 comments

Cite this review

Pith. "Pith review of From perception to control: an autonomous driving system for a formula student driverless car." pith.science (2026). https://pith.science/paper/IOCUCMP7

@misc{pith2026190900119,
  author       = {Pith},
  title        = {Pith review of: From perception to control: an autonomous driving system for a formula student driverless car},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IOCUCMP7}},
  note         = {Machine review of arXiv:1909.00119}
}

read the original abstract

This paper introduces the autonomous system of the "Smart Shark II" which won the Formula Student Autonomous China (FSAC) Competition in 2018. In this competition, an autonomous racecar is required to complete autonomously two laps of unknown track. In this paper, the author presents the self-driving software structure of this racecar which ensure high vehicle speed and safety. The key components ensure a stable driving of the racecar, LiDAR-based and Vision-based cone detection provide a redundant perception; the EKF-based localization offers high accuracy and high frequency state estimation; perception results are accumulated in time and space by occupancy grid map. After getting the trajectory, a model predictive control algorithm is used to optimize in both longitudinal and lateral control of the racecar. Finally, the performance of an experiment based on real-world data is shown.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

8 extracted references · 8 canonical work pages

[1]

An Open Approach to Autonomous V ehicles

S. Kato, E. Takeuchi, Y . Ishiguro, Y . Ninomiya, K. Takeda, and T. Hamada. “An Open Approach to Autonomous V ehicles”. IEEE Micro, V ol. 35, No. 6, pp. 60-69, 2015

work page 2015
[2]

Self-Driving Cars: A Survey

Claudine Badue, Rânik Guidolini, Raphael V . Carneiro. “Self-Driving Cars: A Survey”. 2019

work page 2019
[3]

Design of an Autonomous Racecar: Perception, State Estimation and System Integration

Miguel I. V alls, Hubertus F.C. Hendrikx, Victor J.F. Reijgwart. “Design of an Autonomous Racecar: Perception, State Estimation and System Integration”. 2018 IEEE International Conference on Robotics and Automation. May 21-25, 2018

work page 2018
[4]

Design of an Autonomous Race Car for the Formula Student Driverless (FSD)

Marcel Zeilinger 1, Raphael Hauk 1, Markus Bader. “Design of an Autonomous Race Car for the Formula Student Driverless (FSD)”. 2017

work page 2017
[5]

LOAM : LiDAR Odometry and Mapping in real-time

Zhang, J., and S. Singh. "LOAM : LiDAR Odometry and Mapping in real-time." 2014

work page 2014
[6]

Fast segmentation of 3D point clouds for ground vehicles

M. Himmelsbach, F. v. Hundelshausen, and H. -. Wuensche. “Fast segmentation of 3D point clouds for ground vehicles”. In: 2010 IEEE Intelligent V ehicles Symposium. June 2010, pp. 560–565

work page 2010
[7]

YOLOv3: An Incremental Improvement

Joseph Redmon, Ali Farhadi. “YOLOv3: An Incremental Improvement”. 2018

work page 2018
[8]

Autonomous Driving System Design for Formula Student Driverless Racecar

hanqing TIAN, Jun NI, Jibin HU. “Autonomous Driving System Design for Formula Student Driverless Racecar”. 2018 IEEE Intelligent V ehicles Symposium. June 26-30, 2018

work page 2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

An Open Approach to Autonomous V ehicles

S. Kato, E. Takeuchi, Y . Ishiguro, Y . Ninomiya, K. Takeda, and T. Hamada. “An Open Approach to Autonomous V ehicles”. IEEE Micro, V ol. 35, No. 6, pp. 60-69, 2015

work page 2015

[2] [2]

Self-Driving Cars: A Survey

Claudine Badue, Rânik Guidolini, Raphael V . Carneiro. “Self-Driving Cars: A Survey”. 2019

work page 2019

[3] [3]

Design of an Autonomous Racecar: Perception, State Estimation and System Integration

Miguel I. V alls, Hubertus F.C. Hendrikx, Victor J.F. Reijgwart. “Design of an Autonomous Racecar: Perception, State Estimation and System Integration”. 2018 IEEE International Conference on Robotics and Automation. May 21-25, 2018

work page 2018

[4] [4]

Design of an Autonomous Race Car for the Formula Student Driverless (FSD)

Marcel Zeilinger 1, Raphael Hauk 1, Markus Bader. “Design of an Autonomous Race Car for the Formula Student Driverless (FSD)”. 2017

work page 2017

[5] [5]

LOAM : LiDAR Odometry and Mapping in real-time

Zhang, J., and S. Singh. "LOAM : LiDAR Odometry and Mapping in real-time." 2014

work page 2014

[6] [6]

Fast segmentation of 3D point clouds for ground vehicles

M. Himmelsbach, F. v. Hundelshausen, and H. -. Wuensche. “Fast segmentation of 3D point clouds for ground vehicles”. In: 2010 IEEE Intelligent V ehicles Symposium. June 2010, pp. 560–565

work page 2010

[7] [7]

YOLOv3: An Incremental Improvement

Joseph Redmon, Ali Farhadi. “YOLOv3: An Incremental Improvement”. 2018

work page 2018

[8] [8]

Autonomous Driving System Design for Formula Student Driverless Racecar

hanqing TIAN, Jun NI, Jibin HU. “Autonomous Driving System Design for Formula Student Driverless Racecar”. 2018 IEEE Intelligent V ehicles Symposium. June 26-30, 2018

work page 2018