A Low-Cost, Open Platform for End-to-End Autonomous Driving on a Miniature Ackermann Vehicle
Gustavo Claudio Karl Couto, Eric Aislan Antonelo, Gabriel George Zipperer
cs.LG · cs.AI · cs.RO
Quick verdict
Claims a low-cost open mini-Ackermann sim-to-real platform; sim-augmented BC policy completes all four routes, but physical evidence rests on n=1–3 runs.
Summary
The paper introduces a low-cost, open experimental platform for end-to-end autonomous driving research on a miniature Ackermann vehicle: a physical car (ESP32-S3 micro-ROS firmware, 100 Hz per-wheel PID, UDP camera link), a printed 3.2×2.8 m urban track, teleoperation data-collection tools, map registration for cross-track-error (CTE) metrics, and a Webots digital twin. As a baseline it trains command-conditioned behavior cloning: a 94,882-parameter CNN mapping a camera image plus a discrete navigation command to steering and speed. Three fully autonomous runs at 0.15 m/s reach 6.1±1.2 cm mean CTE versus 4.7±1.8 cm for human teleoperation, but fail the left-turn route. A ~1.0M-parameter variant trained on real data alone completes fewer routes; only when trained with Webots-generated synthetic laps (43,581 frames) translated by a domain-randomized U-Net does it complete all four routes (~5.1 cm overall). In the twin, widening camera field of view from 58° to 120° cuts mean CTE from 35.6 to 3.3 cm. Limitations, largely acknowledged by the authors: very few physical runs (n=1–3 per condition), manual trajectory registration, and ablations that confound capacity with input resolution. Firmware, host software and checkpoints are released under the MIT license, making this a credible community testbed.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The open platform (miniature Ackermann vehicle, printed urban track, data-collection and map-registration tools, Webots digital twin) supports quantitative, map-referenced closed-loop evaluation of end-to-end driving policies.
- Command-conditioned behavior cloning with a 94,882-parameter CNN drives the physical track at 0.15 m/s with 6.1±1.2 cm mean cross-track error, close to the 4.7±1.8 cm of human teleoperation demonstrations.
- Only the ~1.0M-parameter policy trained on real data plus 43,581 sim-to-real-translated Webots frames completes all four track routes; the same network on real data alone and the compact baseline complete fewer.
- Camera field of view is a dominant factor in the digital twin: widening from 58° to 120° reduces mean cross-track error from 35.6 to 3.3 cm.
Strengths
Rare end-to-end systems transparency
Section 5.1 documents the full control stack with numbers a sim-to-real researcher actually needs: 100 Hz per-wheel PID with stated gains, camera transport latency (median 27 ms), capture-to-actuation latency (median 59.5 ms, p95 118.7 ms), 2.5 ms policy inference at 10 Hz, and a 500 ms command watchdog.
Genuine open release on commodity hardware
Footnote 1 commits to an MIT-licensed release of firmware, host software (operator node, inference, camera bridge, EKF, calibration) and the deployed checkpoint on an off-the-shelf Yahboom R2L-class chassis — replication is realistic, not aspirational.
Clean single-factor ablation in the twin
Human-demonstration reference band
Honest limitations section
Weaknesses
Tiny physical samples behind the headline claim
Confounded architecture ablation
Manual trajectory registration unquantified
Asymmetric checkpoint selection
Severe command imbalance unexplored
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Structured Review: A Low-Cost, Open Platform for End-to-End Autonomous Driving on a Miniature Ackermann Vehicle
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How close is the learned driving policy to a human driver in this paper?
On the printed miniature track, the compact behavior-cloning policy achieved a mean cross-track error of 6.1±1.2 cm over three fully autonomous runs at 0.15 m/s, compared with 4.7±1.8 cm for fully localized human teleoperation sessions, both measured after registering trajectories to a rectified map. The policy failed the left-turn route, which only a larger, synthetically augmented policy completed.
Is the code, hardware design, or dataset from this paper openly available?
Yes. The authors state that the vehicle firmware (ESP32-S3 micro-ROS), host software (operator node, policy inference, camera bridge, EKF, calibration files), and the deployed model checkpoint are released under the MIT license at github.com/gustavokcouto/teledriving-in-miniature. The hardware is an off-the-shelf Yahboom ROSMASTER R2L-class Ackermann chassis plus a printed 3.2×2.8 m road mat.
Why does synthetic data from the simulator matter in this paper?
A larger (~1.0M parameter) policy trained only on real teleoperation demonstrations completed just two of the four track routes. Adding 43,581 Webots-simulated frames, converted to look like real camera images by a domain-randomized U-Net translator, was the only tested configuration that completed all four routes in closed loop, with an overall mean cross-track error of about 5.1 cm.
What did the camera field-of-view experiment show?
In the Webots digital twin with exact ground truth, the calibrated narrow 58° lens produced a mean cross-track error of 35.6±12.1 cm, while widening to 120° reduced it to 3.3±0.2 cm. The authors conclude the camera's field of view is a major limiting factor for route following on this platform.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figure images could not be inspected; judgments about Fig. 7 (n counts, per-route CTE labels) rest on the extracted panel text, which appeared complete and internally consistent with Sections 5.6–5.8.
- The abstract string in the supplied metadata is cut mid-sentence ('These results estab…'), but the full abstract and complete body through the reference list are present in the extracted text, so no content judgments were affected.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.