Pith. sign in

REVIEW 5 major objections 5 minor 54 references

Safety-Prioritized, Reinforcement Learning-Enabled Traffic Flow Optimization in a 3D City-Wide Simulation Environment

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a safety-prioritized PPO traffic-signal controller, trained in a physics-based 3D city simulation, cuts serious collisions by 75%, vehicle-vehicle collisions by 79%, and vehicle-non-vehicle collisions by 96% while…

desk verdict A real feasibility demo of Unity-based city-scale RL traffic control, but the headline distance, fuel, and carbon numbers don't survive the baseline survival confound. read the letter →

arxiv 2506.03161 v1 pith:G43B4AUC submitted 2025-05-23 cs.LG cs.AIcs.RO

Mira Nuthakki This is my paper · ORCID
classification cs.LGcs.AIcs.RO
keywords reinforcementlearningtrafficsimulationvehiclesafetyflowoptimizationdeepintelligenttransportationsystemsproximalpolicy3Dcity
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

The paper is trying to establish that a reinforcement-learning agent trained with a safety-prioritized reward can manage city-wide traffic signals inside a physics-based 3D simulation, reducing collisions without sacrificing throughput. The author builds a game-engine city with 46 signalized intersections and up to 879 vehicles in which collisions emerge from physical interactions instead of being scripted, then trains a proximal policy optimization (PPO) controller that adjusts green-light durations and speed limits. In 600-second tests, the trained controller reduces serious collisions from 424 to 106, vehicle-vehicle collisions from 5082 to 1046, and vehicle-non-vehicle collisions from 1576 to 60, while total distance traveled rises from 338.62 to 1509.23 units. The paper also reports a 39.33% fuel-efficiency gain and an 88% cut in per-vehicle carbon emissions. If these results hold, they would show that 3D game-engine simulations can serve as testbeds for safety-prioritized, collision-modeled traffic-light control.

What carries the argument

The load-bearing mechanism is a custom reward function layered on top of a centralized proximal policy optimization (PPO) agent and a physics-enabled collision environment. The reward continuously penalizes stopped time at $-10^{-5}$ per step, rewards distance traveled at $+10^{-8}$, and rewards time spent in the 25-30 units/s speed band at $+10^{-5}$; discretely, it gives $+0.01$ for passing an intersection, $-1$ for a serious collision, and $-0.01$ for any vehicle collision. Vehicles are governed by wheel-collider physics, navigate a waypoint-based road network, and are removed from the simulation after being stopped for 60 seconds in a collision, which the paper treats as towing after a serious crash. The agent observes sampled vehicle and traffic-light positions plus current signal phase and green duration, and its action modifies green-phase duration and, secondarily, vehicle speed limits, with green-light decisions made at 60-second intervals.

What would settle it

Re-run the same 600-second comparison with a matched control that respawns replacement vehicles whenever a collision removes one, or report per-survivor distance and per-survivor stop time; if the PPO agent's 345% distance advantage and 75% collision reduction do not survive that adjustment, the central claim would be falsified.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that a single centralized PPO agent can learn green-light timings that prevent collisions and improve flow in a heterogeneous 3D urban network. The paper reports that over three random-seed test runs the agent cut serious collisions by 75% (424 to 106), total vehicle-vehicle collisions by 79% (5082 to 1046), and vehicle-non-vehicle collisions by 96% (1576 to 60), while increasing average distance traveled per car by 345% (338.62 to 1509.23 units). The author interprets the simultaneous rise in stopped time (2.18 to 4.59 minutes per car) as a consequence of vehicle survival: the baseline removes many vehicles after serious collisions, so fewer remain to produce late-episode congestion. The improvements are attributed mainly to optimized traffic-light signal timing rather than speed-limit adjustment, and the model retained a smaller but significant benefit (21% distance improvement) when transferred to a larger 67-intersection, 1896-vehicle city without retraining. The paper also reports a 39.33% improvement in per-vehicle fuel efficiency and an 88% reduction in per-vehicle carbon emissions.

Load-bearing premise

The load-bearing premise is that the no-RL baseline is a fair control; but because vehicles in the baseline are removed after serious collisions, its lower distance and congestion partly reflect a smaller surviving vehicle population rather than worse traffic control.

Editorial extensions

If this is right

  • Traffic-signal green-phase duration alone is a sufficient control lever for collision reduction in dense signalized networks; the paper found speed-limit adjustments contributed little because intersections are close together in the downtown grid.
  • The trained controller transfers across city scale without retraining, achieving a 21% improvement in distance traveled at 67 intersections with 1896 vehicles (p=0.0073), which suggests the learned policy is at least partly generalizable to denser networks.
  • The physics-enabled simulator produces statistically distinguishable collision outcomes in 10-minute episodes, which makes it possible to train RL agents on realistic collision events instead of relying on rare real-world crashes.
  • The reported environmental benefits, 39.33% higher fuel efficiency and 88% lower carbon emissions per vehicle, coexist with a 345% increase in distance traveled, showing the claimed gains are not achieved by reducing travel.

Reading between the lines

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

  • Because the fuel-efficiency and carbon-emission figures rely on an emissions model the paper does not state or cite, those two environmental numbers should be treated as provisional until the model is published and independently checked.
  • A matched baseline that respawns collision-removed vehicles would let future work separate survival effects from flow improvements; the paper's own explanation of the stopped-time increase indicates such a test is needed before the 345% figure can be read as pure traffic-flow gain.
  • The reward-weight disparity, with a collision costing $-1$ while one stopped-time step costs only $10^{-5}$, suggests that sweeping reward coefficients and measuring the resulting safety-efficiency frontier could quantify how much throughput must be traded for each additional collision prevented.
  • The author's closing suggestion that the same physics-informed conflict-resolution approach could apply to energy grids or communication relays would require rebuilding the vehicle-specific physics and reward models for those domains, so it is best read as a research direction rather than a demonstrated transfer.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper presents a Unity-based 3D urban traffic simulator combining macroscopic and microscopic dynamics, a physics-based collision model, and a proximal policy optimization (PPO) agent that adjusts traffic signal green durations and vehicle speed limits. The authors report that the trained policy reduces serious collisions by 75%, vehicle-vehicle collisions by 79%, and vehicle-non-vehicle collisions by 96%, while increasing total distance traveled by 345% (from 338.62 to 1509.23 units), improving fuel efficiency by 39.33%, and reducing carbon emissions by 88% per vehicle. They argue that these results establish the feasibility of game-engine-based 3D traffic simulation for safety-prioritized, congestion-reducing control aligned with Vision Zero principles.

Significance. If the reported results were sound, the 3D Unity simulation environment with physics-enabled collision modeling and an RL traffic-signal controller would be a useful feasibility demonstration, particularly because the code and data are publicly available and the environment addresses a real gap in collision-focused traffic simulation. However, the main quantitative claims are not supported by the evidence as presented: the distance-traveled increase is confounded by differential vehicle removal, the fuel and carbon numbers rely on an unspecified model, and the collision reductions are direct targets of the learned reward. The paper therefore does not currently establish the claimed traffic-flow or environmental benefits.

major comments (5)
  1. [Results, Model Performance (Figs. 2, 7, 8)] The 345% increase in total and average distance traveled is not a valid measure of traffic-flow improvement because the baseline and model conditions have different surviving vehicle populations. The paper states that in the baseline 'many vehicles are removed early due to their involvement in serious collisions, resulting in fewer remaining vehicles and consequently less congestion in later periods,' and serious-collision vehicles are removed after 60 seconds. Since the trained policy reduces serious collisions, the model retains far more vehicles, mechanically inflating total distance and per-vehicle averages. The argument that stopped time increased only 110% versus 345% for distance does not correct for this population imbalance; no matched survival analysis, per-vehicle conditional metrics, or restriction to vehicles that survive the full episode is provided.
  2. [Results, Model Performance (Figs. 9-10)] The reported 39.33% fuel-efficiency improvement and 88% carbon-emission reduction per vehicle are unsupported because no fuel-consumption model, emission factors, or baseline definition are given. The 'per vehicle' basis is also confounded by the differential survival of vehicles discussed above, since the baseline cohort is enriched with short-lived, low-mileage vehicles. These environmental claims should either be removed or fully specified with equations, input data, and a clear statement of what 'per vehicle' means in the presence of vehicle removal.
  3. [Methods, Reward Structure] The collision reductions (75% serious, 79% vehicle-vehicle, 96% vehicle-non-vehicle) are the direct objective of the learned reward: serious collisions incur a penalty of -1 and all vehicle collisions a penalty of -0.01. Reporting that the trained policy reduces collisions therefore reports that the agent achieved its own optimization target. To support a claim of generalizable safety improvement, the paper would need an independent evaluation (e.g., held-out road networks, varied traffic densities, surrogate safety metrics such as TTC, or a comparison against an RL agent trained without collision penalties) or at minimum an explicit acknowledgment that these counts are the training objective rather than an emergent benefit.
  4. [Results, Model Performance and Statistical Reporting] The statement 'All of the p values were 10^-23 or less except for time stopped which had a p value of 10^-5' is not verifiable: the statistical test is not named, the comparison (per simulation, per vehicle, per time step?) is not defined, and only 3 random seed trials are reported. With n=3, such extreme p-values require explanation. Without details on the test procedure and effect sizes, the statistical significance claims are not meaningful and should be omitted or properly documented.
  5. [Abstract and Results] The abstract claims the model reduces 'total distance travelled by over 3 times the baseline values,' but the results show total distance increased from 338.62 to 1509.23 units, a 345% increase. This is a directionally incorrect description of the reported outcome. The sentence should either say the model increases total distance traveled by over 3 times, or the underlying claim needs reconsideration; as written, it misrepresents the central result.
minor comments (5)
  1. [Supplement Section 3, Fig. 4 caption] The caption says 'PPO in 897 vehicles' while the main text consistently says 879 vehicles; please reconcile the discrepancy.
  2. [Methods, Vehicle Physics and RL Actions] The text uses '20-35 mph' in one place but the simulation uses 'units/sec' elsewhere; clarify the mapping between units and real-world speeds.
  3. [Results, Figs. 2-10] The figures are not described in enough detail in the text: the units, error bars, trial counts, and the exact quantity plotted (per-vehicle versus total) should be stated in the captions or in the main text.
  4. [Results, Model Performance (Fig. 7)] The sentence 'time stopped increased by 110%, from 2.18 min in the baseline to 4.59 min in the model results' should specify that these are per-vehicle averages and state the standard deviations or confidence intervals.
  5. [Methods, RL Training] The 'curiosity reward' mentioned in the training description and Supplement Fig. 4 is not defined in the main text; provide the formula or a reference for the curiosity signal used.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline safety and flow improvements are measured on the same quantities encoded in the reward function, so the central claims reduce to the training objective.

  1. self definitional [Methods – Reinforcement Learning: Reward Structure; Results – Model Performance]
    "Continuous rewards included time stopped* -10^-5, distance traveled* +10^-8, and speed 25-30 timer* +10^-5. Discrete rewards included +0.01 for passing through an intersection, -1 for a serious collision, -0.01 for all vehicle collisions. ... In testing ... the model increased total distance travelled ... by 345%, from 338.62 to 1509.23 ... reduced serious collisions by 75%, from 424 to 106 ... vehicle-vehicle collisions decreased by 79%, from 5082 to 1046 ... time per vehicle in ... 25-30 ... increased by 13.25 times"

    The reward function is composed of the headline outcome variables: -1 per serious collision, -0.01 per vehicle-vehicle collision, +10^-8 per unit distance, and +10^-5 per second in the 25-30 speed bin. The Results section reports improvements in exactly these quantities, so the reported collision reduction, distance increase, and speed-bin time increase are the policy's own training objective. A PPO policy trained to maximize this reward is, by construction, being optimized on these metrics; reporting the trained policy's improvements on them is reporting that the optimization succeeded on its own objective, not a first-principles prediction. The comparison is further confounded by differential vehicle removal in the baseline, which the paper itself admits.

full rationale

The paper has no load-bearing self-citations, and the Unity-based 3D environment, collision physics, PPO/SAC/MA-POCA comparisons, and cross-environment validation on 1896 vehicles provide independent empirical content. The circularity is narrower but central: the reward function literally defines the metrics used to claim success, so the abstract's claim that the model substantially reduces serious collisions, vehicle-vehicle collisions, and distance traveled restates the training objective. This is partial, not total, circularity because the optimization could fail and the cross-environment test (21% distance improvement on a different 67-intersection, 1896-vehicle environment) offers some independent evidence. The paper's own admission that baseline vehicles are 'removed early due to their involvement in serious collisions' is a serious control-validity confound for the distance and fuel/carbon claims, but that is a correctness risk, not circularity. The 39.33% fuel-efficiency and 88% carbon-emission reductions are asserted without a stated fuel or emissions model, which is a missing-support issue rather than a circular one. Overall score 6: some headline 'predictions' reduce to the training objective by construction.

Assumptions & free parameters 10 free parameters · 6 assumptions · 0 invented entities

The paper introduces no new physical entities. Its quantitative claims rest on a large set of hand-tuned parameters (reward weights, collision thresholds, raycast distances) and on strong domain assumptions about the faithfulness of the Unity physics simulation, the validity of the artificially high collision environment, and the correctness of an unstated fuel/emission model. The free parameters directly shape the reported outcomes, particularly the collision reductions, which are explicit reward objectives.

free parameters (10)
  • Reward weight for serious collision = -1
    Hand-chosen to heavily penalize serious collisions; directly shapes the reported collision reduction.
  • Reward weight for vehicle-vehicle collision = -0.01
    Small penalty for any vehicle collision, calibrated during reward shaping.
  • Reward weight for intersection pass-through = +0.01
    Positive reward for passing through an intersection, encouraging flow.
  • Reward weight for distance traveled = +10^-8
    Small positive reward per unit distance, encouraging vehicles to move.
  • Reward weight for time stopped = -10^-5
    Penalty per unit time stopped, shaping congestion behavior.
  • Reward weight for speed 25-30 timer = +10^-5
    Bonus for time spent in target speed range, encouraging free-flow speeds.
  • Serious collision threshold = 30 seconds stopped
    Arbitrary cutoff defining serious collisions; the paper notes this is an operational label.
  • Vehicle removal threshold = 60 seconds stopped
    Vehicles stopped over 60 seconds are removed, simulating towing; this affects survival and distance metrics.
  • Forward raycast distance = 6 units
    Detection range for obstacle avoidance; short range inflates collision frequency, a key driver of results.
  • Side raycast angle and distance = ±37-45 degrees, 2 units
    Angles and distances for side detection, tuned to balance collision frequency and realism.
assumptions (6)
  • domain assumption Unity's PhysX physics engine produces vehicle collisions representative of real-world dynamics.
    The collision model is entirely Unity's physics engine; no validation against real collision data is provided.
  • domain assumption The raycast-based perception system with 6-unit forward range approximates real driver perception and reaction.
    Collision frequency is driven by this artificial limit; real drivers have longer perception distances.
  • ad hoc to paper Time stopped is a valid proxy for collision severity.
    Serious collisions are defined as vehicles stopped over 30 seconds; this cutoff has no physical, medical, or legal basis.
  • domain assumption The deliberately high collision environment is representative enough to test safety-prioritized control.
    The paper states a low collision environment could not achieve statistical significance (Supplement Fig 15), so the environment was made collision-heavy; results may not transfer to realistic rates.
  • domain assumption The fuel and carbon emission computation is valid.
    No methodology, model, or citation for fuel efficiency or carbon emission calculations is provided; the conclusions rely on an invisible computation.
  • domain assumption PPO with the listed hyperparameters learns a policy that generalizes within and beyond the training city.
    Only one training environment is used, with 50 randomly sampled vehicles per frame; cross-environment validation shows reduced improvement (21% distance increase) in the larger city.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Safety-Prioritized, Reinforcement Learning-Enabled Traffic Flow Optimization in a 3D City-Wide Simulation Environment." pith.science (2026). https://pith.science/paper/G43B4AUC

@misc{pith2026250603161,
  author       = {Pith},
  title        = {Pith review of: Safety-Prioritized, Reinforcement Learning-Enabled Traffic Flow Optimization in a 3D City-Wide Simulation Environment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G43B4AUC}},
  note         = {Machine review of arXiv:2506.03161}
}
read the original abstract

Traffic congestion and collisions represent significant economic, environmental, and social challenges worldwide. Traditional traffic management approaches have shown limited success in addressing these complex, dynamic problems. To address the current research gaps, three potential tools are developed: a comprehensive 3D city-wide simulation environment that integrates both macroscopic and microscopic traffic dynamics; a collision model; and a reinforcement learning framework with custom reward functions prioritizing safety over efficiency. Unity game engine-based simulation is used for direct collision modeling. A custom reward enabled reinforcement learning method, proximal policy optimization (PPO) model, yields substantial improvements over baseline results, reducing the number of serious collisions, number of vehicle-vehicle collisions, and total distance travelled by over 3 times the baseline values. The model also improves fuel efficiency by 39% and reduces carbon emissions by 88%. Results establish feasibility for city-wide 3D traffic simulation applications incorporating the vision-zero safety principles of the Department of Transportation, including physics-informed, adaptable, realistic collision modeling, as well as appropriate reward modeling for real-world traffic signal light control towards reducing collisions, optimizing traffic flow and reducing greenhouse emissions.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 43 canonical work pages

  1. [1]

    Number of cars in U.S

    "Number of cars in U.S." (2017). Statista. https://www.statista.com/statistics/183505/number of-vehicles-in-the-united-states-since-1990/

    work page 2017

  2. [2]

    How Many Cars Are There in the World in 2021? Stats by Country

    "How Many Cars Are There in the World in 2021? Stats by Country" (2021, June 24). Hedges & Company. https://hedgescompany.com/blog/2021/06/how-many-cars-are-there-in-the-world/

    work page 2021

  3. [3]

    Stein, J. (1990). *New Cars for Better Future: Driving Us Crazy*. Earthgreen

    work page 1990

  4. [4]

    (2019, March 7)

    Fleming, S. (2019, March 7). *Traffic congestion cost the US economy nearly $87 billion in 2018*. World Economic Forum. https://www.weforum.org/agenda/2019/03/traffic-congestion cost-the-us-economy-nearly-87-billion-in-2018/

    work page 2019

  5. [5]

    Urban mobility in the EU Audit preview Information on an upcoming audit

    "Urban mobility in the EU Audit preview Information on an upcoming audit" (2019). https://www.eca.europa.eu/lists/ecadocuments/ap19_07/ap_urban_mobility_en.pdf

    work page 2019

  6. [6]

    Mikkonen, S., Laine, M., Mäkelä, H., Gregow, H., Tuomenvirta, H., Lahtinen, M., & Laaksonen, A. (2015). Trends in the average temperature in Finland, 1847–2013. *Stochastic Environmental Research and Risk Assessment, 29*, 1521–1529

    work page 2015

  7. [7]

    A., & Zaloshnja, T

    Blincoe, L., Ted, R., Miller, E., Zaloshnja, B. A., & Zaloshnja, T. R. (2015). *The Economic and Societal Impact Of Motor Vehicle Crashes, 2010 (Revised)* (Vol. 2277, pp. 8–728). https://crashstats.nhtsa.dot.gov/Api/Public/ViewPublication/812013

    work page 2015

  8. [8]

    NHTSA Releases 2020 Traffic Crash Data

    "NHTSA Releases 2020 Traffic Crash Data" (2020). US Department of Transportation. https://www.transportation.gov/briefing-room/nhtsa-releases-2020-traffic-crash-data

    work page 2020

Show all 54 references
  1. [9]

    Summary of Motor Vehicle Crashes: 2022 Data 1 Summary of Motor Vehicle Traffic Crashes Key Findings

    "Summary of Motor Vehicle Crashes: 2022 Data 1 Summary of Motor Vehicle Traffic Crashes Key Findings" (2024)

    2024

  2. [10]

    Road traffic injuries

    "Road traffic injuries" (2023, December 13). World Health Organization. https://www.who.int/news-room/fact-sheets/detail/road-traffic-injuries

    2023

  3. [11]

    68% of the World Population Projected to Live in Urban Areas by 2050, Says UN

    "68% of the World Population Projected to Live in Urban Areas by 2050, Says UN" (2018, May 16). United Nations Department of Economic and Social Affairs; United Nations. https://www.un.org/development/desa/en/news/population/2018-revision-of-world-urbanization prospects.html

    2018

  4. [12]

    A., Ramadan Mazrekaj, & Hoxha, G

    Rrecaj, A. A., Ramadan Mazrekaj, & Hoxha, G. (2024). A Modified Webster Model for a Useful Traffic Signal Timing Plan. *Lecture Notes on Multidisciplinary Industrial Engineering*, 108–120. https://doi.org/10.1007/978-3-031-48933-4_11

    2024 doi

  5. [13]

    Kumar, R. (2011). Vehicle-Actuated Signal Controller For Heterogeneous Traffic Having Limited Lane Discipline. *Ite Journal*, *May 2011*, 44–53. https://www.researchgate.net/publication/259533291_Vehicle Actuated_Signal_Controller_For_Heterogeneous_Traffic_Having_Limited_Lane_...

    2011

  6. [14]

    Demiral, C., & Celikoglu, H. B. (2011). Application of ALINEA ramp control algorithm to freeway traffic flow on approaches to Bosphorus strait crossing bridges. *Procedia - Social and Behavioral Sciences*, *20*, 364–371. https://doi.org/10.1016/j.sbspro.2011.08.042

    2011 doi

  7. [15]

    Www.scats.nsw.gov.au

    "SCATS" (n.d.). Www.scats.nsw.gov.au. https://www.scats.nsw.gov.au/home

  8. [16]

    Lewandowski, M., Płaczek, B., & Bernas, M. (2017). Self-organizing Traffic Signal Control with Prioritization Strategy Aided by Vehicular Sensor Network. *Lecture Notes in Computer Science*, 536–547. https://doi.org/10.1007/978-3-319-59105-6_46

    2017 doi

  9. [17]

    Han, Y., Hegyi, A., Zhang, L., He, Z., Chung, E., & Liu, P. (2022). A new reinforcement learning-based variable speed limit control approach to improve traffic efficiency against freeway jam waves. *Transportation Research Part C: Emerging Technologies*, *144*, 103900– 103900....

    2022

  10. [18]

    CityFlow

    "CityFlow" (2024, July 23). GitHub. https://github.com/cityflow-project/CityFlow

    2024

  11. [19]

    Sun, S., Chen, J., & Sun, J. (2019). Traffic congestion prediction based on GPS trajectory data. *International Journal of Distributed Sensor Networks*, *15*(5), 155014771984744. https://doi.org/10.1177/1550147719847440

    2019 doi

  12. [20]

    K., & Talebpour, A

    Hosseini, M. K., & Talebpour, A. (2019). Traffic Prediction using Time-Space Diagram: A Convolutional Neural Network Approach. *Transportation Research Record: Journal of the Transportation Research Board*, 036119811984129. https://doi.org/10.1177/0361198119841291

    2019 doi

  13. [21]

    J., Arjona, J., Linares, M., & Casanovas-Garcia, J

    Vázquez, J. J., Arjona, J., Linares, M., & Casanovas-Garcia, J. (2020). A Comparison of Deep Learning Methods for Urban Traffic Forecasting using Floating Car Data. *Transportation Research Procedia*, *47*, 195–202. https://doi.org/10.1016/j.trpro.2020.03.079

    2020 doi

  14. [22]

    (2018, May 30)

    Beall, A. (2018, May 30). *In China, Alibaba's data-hungry AI is controlling (and watching) cities*. Wired. https://www.wired.com/story/alibaba-city-brain-artificial-intelligence-china-kuala lumpur/

    2018

  15. [23]

    A., Abdel-Hamid, Y., & Hefny, H

    Sayed, S. A., Abdel-Hamid, Y., & Hefny, H. A. (2023). Artificial intelligence-based traffic flow prediction: a comprehensive review. *Journal of Electrical Systems and Information Technology*, *10*(1). https://doi.org/10.1186/s43067-023-00081-6

    2023 doi

  16. [24]

    Siri, S., Pasquale, C., Sacone, S., & Ferrara, A. (2021). Freeway traffic control: A survey. *Automatica*, *130*, 109655. https://doi.org/10.1016/j.automatica.2021.109655

    2021

  17. [25]

    Xi, Y.-G., Li, D.-W., & Lin, S. (2013). Model Predictive Control — Status and Challenges. *Acta Automatica Sinica*, *39*(3), 222–236. https://doi.org/10.1016/S1874-1029(13)60024-5

    2013 doi

  18. [26]

    Han, Y., Ramezani, M., Hegyi, A., & Yuan, Y. (2020). Hierarchical ramp metering in freeways: An aggregated modeling and control approach. *Transportation Research Part C Emerging Technologies*, *110*, 1–19. https://doi.org/10.1016/j.trc.2019.09.023

    2020 doi

  19. [27]

    G., Machemehl, R., & Perrine, K

    Liu, H., Claudel, C. G., Machemehl, R., & Perrine, K. A. (2021). A Robust Traffic Control Model Considering Uncertainties in Turning Ratios. *IEEE Transactions on Intelligent Transportation Systems*, *23*(7), 6539–6555. https://doi.org/10.1109/tits.2021.3058315

    2021

  20. [28]

    Tettamanti, T., Luspay, T., Kulcsar, B., Peni, T., & Varga, I. (2014). Robust Control for Urban Road Traffic Networks. *IEEE Transactions on Intelligent Transportation Systems*, *15*(1), 385–398. https://doi.org/10.1109/tits.2013.2281666

    2014

  21. [29]

    Han, Y., Wang, M., & Leclercq, L. (2023). Leveraging reinforcement learning for dynamic traffic control: A survey and challenges for field implementation. *Communications in Transportation Research*, *3*, 100104. https://doi.org/10.1016/j.commtr.2023.100104

    2023

  22. [30]

    Abdulhai, B., Pringle, R., & Karakoulas, G. J. (2003). Reinforcement Learning for True Adaptive Traffic Signal Control. *Journal of Transportation Engineering*, *129*(3), 278–285. https://doi.org/10.1061/(asce)0733-947x(2003)129:3(278)

    2003 doi

  23. [31]

    Aradi, S. (2020). Survey of Deep Reinforcement Learning for Motion Planning of Autonomous Vehicles. *IEEE Transactions on Intelligent Transportation Systems*, 1–20. https://doi.org/10.1109/tits.2020.3024655

    2020

  24. [32]

    Belletti, F., Haziza, D., Gomes, G., & Bayen, A. M. (2018). Expert Level Control of Ramp Metering Based on Multi-Task Deep Reinforcement Learning. *IEEE Transactions on Intelligent Transportation Systems*, *19*(4), 1198–1207. https://doi.org/10.1109/tits.2017.2725912

    2018

  25. [33]

    Li, Z., Liu, P., Xu, C., Duan, H., & Wang, W. (2017). Reinforcement Learning-Based Variable Speed Limit Control Strategy to Reduce Traffic Congestion at Freeway Recurrent Bottlenecks. *IEEE Transactions on Intelligent Transportation Systems*, *18*(11), 3204–3217. https://doi.o...

    2017

  26. [34]

    El-Tantawy, S., Abdulhai, B., & Abdelgawad, H. (2013). Multiagent Reinforcement Learning for Integrated Network of Adaptive Traffic Signal Controllers (MARLIN-ATSC): Methodology and Large-Scale Application on Downtown Toronto. *IEEE Transactions on Intelligent Transportation S...

    2013

  27. [35]

    Chen, C., Wei, H., Xu, N., Zheng, G., Yang, M., Xiong, Y., Xu, K., & Li, Z. (2020). Toward A Thousand Lights: Decentralized Deep Reinforcement Learning for Large-Scale Traffic Signal Control. *Proceedings of the AAAI Conference on Artificial Intelligence*, *34*(04), 3414–3421....

    2020 doi

  28. [36]

    Haydari, A., & Yilmaz, Y. (2020). Deep Reinforcement Learning for Intelligent Transportation Systems: A Survey. *IEEE Transactions on Intelligent Transportation Systems*, 1–22. https://doi.org/10.1109/tits.2020.3008612

    2020

  29. [37]

    Xiao, Y., Liu, J., Wu, J., & Ansari, N. (2021). Leveraging Deep Reinforcement Learning for Traffic Engineering: A Survey. *IEEE Communications Surveys & Tutorials*, *23*(4), 2064–

    2021

  30. [38]

    Zeng, J., Yu, C., Yang, X., Ao, W., Hao, Q., Yuan, J., Li, Y., Wang, Y., & Yang, H. (2024). *CityLight: A Universal Model for Coordinated Traffic Signal Control in City-scale Heterogeneous Intersections*. ArXiv.org. https://arxiv.org/abs/2406.02126

    2024 arXiv

  31. [39]

    Traffic Simulations Software: A Comparison of SUMO, PTV Vissim

    "Traffic Simulations Software: A Comparison of SUMO, PTV Vissim" (n.d.). Think Transportation. https://thinktransportation.net/traffic-simulations-software-a-comparison-of sumo-ptv-vissim-aimsun-and-cube/

  32. [40]

    Ni, D. (2020). Limitations of current traffic models and strategies to address them. *Simulation Modelling Practice and Theory*, *104*, 102137. https://doi.org/10.1016/j.simpat.2020.102137

    2020

  33. [41]

    (2021, March 22)

    Olmez, S. (2021, March 22). *3D Urban Traffic Simulator (ABM) in Unity*. CoMSES Network. https://www.comses.net/codebases/32e7be8c-b05c-46b2-9b5f 73c4d273ca59/releases/1.1.0/

    2021

  34. [42]

    Shi, X., Yang, S., & Ye, Z. (2023). Development of a Unity–VISSIM Co-Simulation Platform to Study Interactive Driving Behavior. *Systems*, *11*(6), 269. https://doi.org/10.3390/systems11060269

    2023 doi

  35. [43]

    (2019, August 20)

    Jin, Z., Swedish, T., & Raskar, R. (2019, August 20). *3D Traffic Simulation for Autonomous Vehicles in Unity and Python*. ArXiv.org. https://doi.org/10.48550/arXiv.1810.12552

    2019 doi arXiv

  36. [44]

    Astarita, V., Caliendo, C., Giofrè, V. P. & Russo, I. Surrogate Safety Measures from Traffic Simulation: Validation of Safety Indicators with Intersection Traffic Crash Data. Sustainability 12, 6974 (2020). SUPPLEMENT Video: Simulated 3D city-wide environment comparison with 2...

    2020

  37. [46]

    Very Small: A minimal urban area with a single central block at (0,0,0) and a distance center parameter of 150 units

  38. [47]

    Small: A moderate sized compact environment with a distance center of 200 units, featuring two large blocks positioned either at [(0,0,0), (0,0,300)] or [(-150,0,150), (150,0,150)] based on random selection

  39. [48]

    Medium: A mid-sized environment with a distance center of 300 units, containing four large blocks with positions determined by one of four different layout patterns selected randomly

  40. [49]

    Large: An extensive environment with a distance center of 350 units, containing six large blocks with positions determined by one of four layout configurations. Building Generation System: Buildings were categorized into nine distinct groups based on their position and functio...

  41. [50]

    Spatial Orientation: Each waypoint was represented by a solid green sphere (1 unit radius) surrounded by a wire sphere (2 unit radius), making them easily identifiable in the scene view

  42. [51]

    Path Connectivity: Green lines connected sequential waypoints, visualizing the complete vehicle path and highlighting any gaps or discontinuities

  43. [52]

    Debugging Navigation: The visualization allowed developers to visually trace vehicle paths through the environment, identify problematic intersections, and verify proper waypoint positioning and orientation

  44. [53]

    Intersection Validation: By observing the terminal waypoints at intersections, developers could verify that proper connectivity was established between different path segments

  45. [54]

    Editor-Only Overhead: Since Gizmos only render in the Unity Editor, they added no performance overhead to the runtime simulation. The Gizmo system was vital during the development process, as it made the otherwise invisible navigation infrastructure visually apparent, allowing...

    2023

  46. [2097]

    https://doi.org/10.1109/comst.2021.3102580

    2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.