CoMoCAVs: Cohesive Decision-Guided Motion Planning for Connected and Autonomous Vehicles with Multi-Policy Reinforcement Learning

Pan Hu

REVIEW 4 major objections 5 minor 35 references

CoMoCAVs: Cohesive Decision-Guided Motion Planning for Connected and Autonomous Vehicles with Multi-Policy Reinforcement Learning

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read CDGMP, the proposed Mixture-of-Experts-style framework, claims that gating a pool of lane-specialist motion planners behind a lane-selection policy makes highway lane changes efficient, revisable, and safer in simulation.

desk verdict A coherent modular RL architecture for highway lane changing that needs baselines and a decision-consistency analysis before its safety/efficiency claims can be taken seriously. read the letter →

arxiv 2507.14903 v1 pith:KEILPGRI submitted 2025-07-20 cs.RO

Pan Hu This is my paper · ORCID

classification cs.RO

keywords connectedautonomousvehicleslane-changedecision-makingmotionplanningmixtureofexpertsmulti-policyreinforcementlearninghierarchicalhighwaysimulationcollisionavoidance

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 tries to establish that highway lane-changing for connected autonomous vehicles can be solved by splitting the task into a high-level lane-selection policy and a pool of lane-specialist motion policies, coordinated by a gate that activates one expert per timestep. The framework, called CDGMP, claims this Mixture-of-Experts-style decomposition improves computational efficiency because only the relevant sub-policy runs during inference, and improves safety because each expert is specialised to one lane. It also claims flexibility: the lane decision is re-evaluated and can be revised at every timestep, unlike prior hierarchical designs where a lane change is irrevocable once initiated. In a two-lane circular highway simulation with slow-moving vehicles, the best configuration (Transformer backbone, safety threshold 1.0) averages 2.67 collisions over 2,000 steps and reaches zero collisions in its best seed, which is the empirical support for the claim. A sympathetic reader would care because most RL lane-change benchmarks ignore the motion-control stage or treat lane changes as instantaneous jumps.

What carries the argument

The load-bearing mechanism is an MoE-inspired gate: a discrete high-level decision policy acts as the router, and each lane of the highway has its own trained expert policy as a specialist. At every timestep the gate chooses a target lane, the corresponding expert's parameters are activated to produce speed $a_v$ and steering $a_\delta$, and the lanelet-based reference path is switched to the centerline of the target lane to encode the intention for the low-level controller. This routing decomposes a continuous control problem into a small number of lane-reaching subproblems, so the inference cost is that of one small expert rather than a monolithic network, and each expert's reward function can be shaped for a single lane in isolation.

What would settle it

Re-run the trained framework in a scenario where a slow vehicle enters the target lane after the gate commits to it but before the expert completes the manoeuvre; if no decision-revision or safety mechanism exists, collisions should increase sharply. A second concrete check is to record, across all seeds, the distribution of the time between a gate decision and expert lane completion: if that distribution's tail exceeds the roughly ten-step consistency window, the composition claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a tightly coupled decision-motion architecture can make lane selection and trajectory execution cohere without a monolithic end-to-end network. The high-level policy consumes a 33-dimensional bird's-eye-view vector and outputs a discrete target-lane index; that index selects a low-level expert policy, each a small MLP consuming a 57-dimensional vector and outputting speed and steering commands, and updates the lanelet-based reference path to the chosen lane's centerline. The two levels are trained sequentially with PPO—experts first, then the decision policy—and the empirical result is that this composition reliably completes highway lane changes, requiring roughly ten consistent decisions per manoeuvre, while tolerating occasional false signals. The paper presents this as evidence that modular specialisation, not network size or end-to-end learning, is the efficient and safe route to CAV highway autonomy in simulation.

Load-bearing premise

The load-bearing assumption is that the separately trained high-level decision policy and the low-level lane experts compose correctly: the gate may propose a lane, but nothing guarantees the corresponding expert can actually move the vehicle there within the roughly ten steps a lane change takes.

Editorial extensions

If this is right

At inference only one specialist policy is active per timestep, so computational cost scales with a single expert, not with the number of experts in the pool.
New driving scenarios can be added by training a new specialist policy and placing it behind the gate, without retraining the full system.
Because the high-level decision is re-evaluated every timestep, a lane-change commitment can be aborted or revised, in contrast to hierarchical designs whose decisions are fixed once initiated.
The framework is backbone-agnostic: naive MLP, LSTM, and Transformer encoders all work inside it, with the Transformer showing the lowest collision counts in the reported runs.
With a well-tuned safety threshold the best simulated runs reach zero collisions over 2,000 steps, indicating that collision-free highway navigation is attainable within this environment.

Reading between the lines

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

A natural extension, not stated in the paper, is to treat the sequential and independent training of experts and gate as the main open risk: the framework assumes that any decision the gate emits is executable by some expert in time, but no closed-loop or joint training verifies that composition.
The roughly ten-step decision consistency suggests a hidden temporal commitment: even though decisions are revisable each step, a manoeuvre only succeeds if the gate holds for about ten steps, so a formal notion of decision latency could predict when hesitation is harmful.
The same gating pattern could transfer to other high-dimensional decision-and-control tasks, but only where experts are genuinely separable; tasks requiring coordinated simultaneous behaviour across experts would not fit.
Following the paper's own outlook, a concrete testable extension is to constrain the high-level decision set with a safety filter at each timestep, so the gate only proposes lanes that the corresponding expert can actually reach.

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.

Desk Editor's Note

read the letter

Colleague,

The one thing to know: this paper builds a clean two-level RL architecture for highway lane changing—two lane-specialist motion policies gated by a high-level decision policy that can revise its lane choice every timestep—and then spends most of its evidence budget on internal backbone comparisons and a safety-threshold sweep. The architecture is coherent and the MoE-style gating over lane experts is a reasonable building block, but the headline claims about safety guarantees and efficiency are not established by the experiments.

What it does well: the observation schemas and reward tables are specified in enough detail to reimplement, the lanelet-based reference path mechanism is a sensible way to couple decisions to motion, and the paper is honest about limitations (hesitation, close distances, CBF as future work). Running three backbone networks and sweeping the threshold is a legitimate sanity check even if it isn't a comparative evaluation.

The soft spots are the ones you'd expect. No baselines against a monolithic policy or any prior hierarchical RL method, so we learn nothing about whether the decomposition actually helps. The efficiency claim has no inference-cost measurement. The safety result comes from picking the best threshold after the fact, and the metric is collision counts on the training loop. The stress-test note lands: the paper itself says a successful lane change needs about 10 consecutive correct decisions, yet the high-level policy can flip at every timestep, the training reward does not penalize flips, and decision-flip statistics are never reported. The Figure 5 hesitation incident is a direct example of a mid-maneuver flip. So the observed reliability is an emergent property of the gate-expert interaction, not something the architecture guarantees. The conclusion's turn to control barrier functions as future work basically concedes this.

The writing also needs polish (typos, a Wikipedia citation, an explicit 'under preparation for publication' note), which is fine for a thesis draft but adds to the impression of a preliminary report.

Who gets value from this: someone working on modular RL for driving who might borrow the gate-over-lane-experts idea or wants a template for vector-based lanelet observations. As a serious refereed publication in its current form, it isn't there. My recommendation: don't desk-reject the authors' future work, but this version needs baselines, a decision-consistency analysis, and code/data before it earns referee time.

Referee Report

4 major / 5 minor

Summary. The paper proposes Cohesive Decision-Guided Motion Planning (CDGMP), a hierarchical reinforcement-learning architecture for highway lane-changing in connected and autonomous vehicles. A high-level policy selects a target lane at each timestep, and a gating mechanism activates one of two specialized low-level expert policies (left-lane or right-lane tracking) that output speed and steering commands. The two levels are trained sequentially and independently with PPO in the CPM Lab simulator. Evaluation reports total collision counts over 2,000 steps across six environment seeds, comparing Transformer, LSTM, and MLP backbones, and studies the effect of a safety threshold. The best reported average is 2.67 collisions at threshold 1.0, which the paper calls 'nearly collision-free.' The central claims are that the modular MoE-style design improves computational efficiency, maintains safety through specialization, and provides flexibility because decisions can be revised each timestep.

Significance. If validated, the architecture would be a useful contribution: a small set of lane-specialist policies gated by a learned decision network is a plausible and potentially scalable way to couple decision-making with motion planning. The paper is also transparent about its limitations, explicitly acknowledging dangerously close inter-vehicle distances, hesitation behavior, and the plan to incorporate control barrier functions in future work. However, the evidence presented does not establish the central safety and efficiency claims. The evaluation is conducted in the same simulation environment used for training, the only reported metric is collision count, no baseline is compared, and the best result is obtained by tuning the safety threshold against the evaluation metric. The decision-consistency assumption that underlies the safety claim is asserted but never measured. These issues are load-bearing for the paper's headline contributions.

major comments (4)

[§IV-C, Table VII] The evaluation metric is total collision count over 2,000 steps, yet collision avoidance is already the primary objective of both reward schemas (Tables II and IV), and the best reported result (2.67 average collisions at threshold 1.0) is obtained by selecting the safety threshold from the same grid used for evaluation. No baseline is provided against a monolithic policy, a prior hierarchical method, or a single-expert policy, so the observed collision counts cannot be attributed to the CDGMP architecture. To support the central claim, the paper needs either an external benchmark, a counterfactual ablation (e.g., a fixed-rule gate or an always-keep-lane policy), or at least a comparison with an end-to-end PPO policy trained on the same observations.
[§IV-D2, Table VII] The 'nearly collision-free' characterization is not statistically supported. With six seeds, the threshold-1.0 configuration has average collisions 2.67 and standard deviation 3.20, overlapping with the threshold-1.5 configuration (5.83 ± 3.31); no significance test or per-seed breakdown is reported. Because the best threshold is chosen from the evaluated grid, the result may reflect selection bias rather than a stable property of the framework. Reporting per-seed collision counts and confidence intervals, or using a held-out threshold, would address this concern.
[§IV-D1, §III-A, §IV-B] The paper's safety claim depends on the unmeasured assumption that high-level decisions remain stable for the roughly 10-step maneuver horizon. The high-level policy can revise decisions at every timestep and the low-level experts are trained only for lane-reaching and centerline tracking, not for recovering from a reference path switched mid-maneuver. Figure 5 documents a one-timestep decision flip, but no statistic is reported for the frequency of such flips, their correlation with collisions, or the system's recovery behavior. Without such an analysis, 'maintains safety guarantees through modular specialization' is an emergent property that has not been established.
[§I-B, §V] The efficiency claim is not quantified. The paper asserts that activating only one specialized small network improves computational efficiency compared with monolithic models, but no inference-time comparison, latency measurement, parameter count, or FLOPs estimate is given anywhere in Section IV. As written, the efficiency advantage is purely architectural and untested.

minor comments (5)

[§II-C] The expected-return formula is garbled by formatting and is unreadable; please use a proper display equation.
[§IV-B] There is a typo: 'gradudally' should be 'gradually'. Also, the spacing of 'CA Vs' is inconsistent throughout the manuscript.
[§IV-D3] The 'Aha Moment' subsection is informal and would be better integrated into a quantitative analysis of decision consistency and recovery behavior, which is currently missing.
[§IV-A] The description of the lanelet-based reference path mechanism says the reference path is updated to the nearest lanelet in the target lane, but it is not specified how this interacts with the 10-step lane-change completion requirement or with mid-maneuver decision revisions.
[References] Citing Wikipedia for reinforcement learning (Ref. [19]) is not ideal for a journal; consider citing the original textbook or a survey article.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported safety evaluation is an in-simulation RL benchmark, not a quantity constructed from the paper's own inputs or from a load-bearing self-citation chain.

full rationale

The paper contains no derivation chain in which a prediction is equivalent to its inputs by construction. High-level and low-level policies are trained with PPO on explicit reward tables (Tables II and IV), and the reported collision counts over 2,000 simulation steps are the same kind of measure encoded in those rewards; this is standard RL validation, not a circular reduction. The safety-threshold sweep in Table VII is an ablation rather than a fitted parameter renamed as a prediction: all threshold values are reported and no held-out forecast is claimed. The 'around 10 consecutive correct decisions' observation in Section IV-D1 is an operational description of lane-change duration, not a definition that forces the outcome. Self-citations are present but not load-bearing: [28] (SigmaRL, co-authored by the present author) supplies the vector-based environment representation, and [30]-[32] are cited only as future work on control barrier functions. No uniqueness theorem or ansatz is imported from the authors' prior work, and no known result is merely renamed. The paper's own conclusion concedes dangerously close overtaking distances and decision hesitation, which weakens the phrasing 'safety guarantees' as a correctness or over-claim concern, but this is a validation gap rather than circularity. Stronger external benchmarks and a counterfactual ablation of the gating mechanism would improve the evidence, but the absence of such tests does not make the derivation circular.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claims rest on reward shaping, sequential training, and simulator fidelity. The tuned safety threshold is an explicit free parameter; reward weights are unspecified. No new physical entities are introduced.

free parameters (2)

safety_threshold = 1.0 (chosen from {0.5, 1.0, 1.5, 2.0})
In the high-level reward, the distance below which a lane is considered risky; the best-performing configuration uses 1.0, selected after evaluating the grid in Table VII.
Reward weights (high-level and low-level) = not reported
Tables II and IV list reward terms but no numeric values; these weights shape the learned behavior and are not published, reducing reproducibility and making the tuned nature of the result harder to assess.

assumptions (4)

domain assumption PPO with the listed hyperparameters converges to a policy that reflects the reward schema.
The paper does not analyze convergence or policy optimality; it assumes training succeeded (Section IV-B).
domain assumption The CPM Lab two-lane loop with slow vehicles is a representative testbed for highway lane-change autonomy.
Section IV-A; the claims of real-world relevance depend on this transfer.
ad hoc to paper Sequential independent training of low-level experts and then the high-level gate yields a cohesive integrated controller.
Section IV-B; this is the key architectural premise, but no joint training or closed-loop composition analysis is provided.
ad hoc to paper The reward engineering (lane-change penalty, safety threshold, centerline deviation penalty) captures the desired driving behavior.
Tables II and IV; the 'aha' hesitation behavior is interpreted as robustness without a formal definition.

how reviews work

0 comments

Cite this review

Pith. "Pith review of CoMoCAVs: Cohesive Decision-Guided Motion Planning for Connected and Autonomous Vehicles with Multi-Policy Reinforcement Learning." pith.science (2026). https://pith.science/paper/KEILPGRI

@misc{pith2026250714903,
  author       = {Pith},
  title        = {Pith review of: CoMoCAVs: Cohesive Decision-Guided Motion Planning for Connected and Autonomous Vehicles with Multi-Policy Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KEILPGRI}},
  note         = {Machine review of arXiv:2507.14903}
}

read the original abstract

Autonomous driving demands reliable and efficient solutions to closely related problems such as decision-making and motion planning. In this work, decision-making refers specifically to highway lane selection, while motion planning involves generating control commands (such as speed and steering) to reach the chosen lane. In the context of Connected Autonomous Vehicles (CAVs), achieving both flexible and safe lane selection alongside precise trajectory execution remains a significant challenge. This paper proposes a framework called Cohesive Decision-Guided Motion Planning (CDGMP), which tightly integrates decision-making and motion planning using a Mixture of Experts (MoE) inspired architecture combined with multi-policy reinforcement learning. By coordinating multiple specialized sub-networks through a gating mechanism, the method decomposes the complex driving task into modular components. Each sub-network focuses on a specific aspect of driving, improving efficiency by activating only the most relevant modules during inference. This design also enhances safety through modular specialization. CDGMP improves the adaptability and robustness of CAVs across diverse traffic scenarios, offering a scalable solution to real-world autonomy challenges. The architectural principles behind CDGMP, especially the use of MoE, also provide a strong foundation for other high-dimensional decision and control tasks. Simulation results (available at https://youtu.be/_-4OXNHV0UY) demonstrate reliable performance in both lane selection and motion planning.

Figures

Figures reproduced from arXiv: 2507.14903 by the authors.

**Figure 1.** Illustration of the proposed framework. MPRL refers to reinforcement learning systems that involve multiple policy networks. While no formal definition of this term yet exists, the prevailing approach in MPRL involves training distinct policies either for multiple objectives or to enhance system robustness [23] [24]. In our work, we specifically employ multiple policies trained to achieve different objectives with… view at source ↗

**Figure 2.** CPM Lab map. IV. EXPERIMENT A. Environmental Setup All experiments are conducted using the map from CPM Lab [27] and largely adhering to the vector-based environmental representations proposed in SigmaRL framework [28]. To accurately replicate highway driving conditions, our implementation specifically utilizes the two outer circular lanes of the CPM Lab map (marked in orange in Figure 2), effectively creating a… view at source ↗

**Figure 3.** Episode reward from training. As shown in Table VI, the Transformer model demonstrates superior collision avoidance performance. It averages 6.67 collisions over 2,000 simulation steps across six different environment seeds, with the best performance reaching as low as 2 collisions. This clearly outperforms both the LSTM and MLP models, which average 7.17 and 8.17 collisions, respectively. TABLE VI: Collisions ove… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Lane changing scenario. free results. The proposed framework is capable of handling occasional false signals but still executes the lane change reliably when it becomes necessary. V. CONCLUSIONS In this work, we presented a novel framework for decisionmaking and motio…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 29 canonical work pages

[1]

Perception, planning, control, and coordination for autonomous vehicles,

S. D. Pendleton, H. Andersen, X. Du, X. Shen, M. Meghjani, Y . H. Eng, D. Rus, and M. H. Ang, “Perception, planning, control, and coordination for autonomous vehicles,” Machines, vol. 5, no. 1, 2017

work page 2017
[2]

A comprehensive review on safe reinforcement learning for au- tonomous vehicle control in dynamic environments,

R. Inamdar, S. K. Sundarr, D. Khandelwal, V . D. Sahu, and N. Katal, “A comprehensive review on safe reinforcement learning for au- tonomous vehicle control in dynamic environments,” e-Prime - Ad- vances in Electrical Engineering, Electronics and Energy , vol. 10, p. 100810, 2024. (a) Timestep 1106. (b) Timestep 1107. (c) Timestep 1108. Fig. 5: Hesitation...

work page 2024
[3]

Review of decision-making and planning approaches in automated driving,

F. Garrido and P. Resende, “Review of decision-making and planning approaches in automated driving,” IEEE Access, vol. 10, pp. 100 348– 100 366, 2022

work page 2022
[4]

Survey on artificial intelligence for vehicles,

J. Li, H. Cheng, H. Guo, and S. Qiu, “Survey on artificial intelligence for vehicles,” Automotive Innovation, vol. 1, pp. 2–14, 2018

work page 2018
[5]

Decision-making technology for autonomous vehicles: Learning-based methods, applications and future outlook,

Q. Liu, X. Li, S. Yuan, and Z. Li, “Decision-making technology for autonomous vehicles: Learning-based methods, applications and future outlook,” in 2021 IEEE International Intelligent Transportation Systems Conference (ITSC) , 2021, pp. 30–37

work page 2021
[6]

Rule-based decision-making system for autonomous vehicles at intersections with mixed traffic environment,

A. Aksjonov and V . Kyrki, “Rule-based decision-making system for autonomous vehicles at intersections with mixed traffic environment,” in 2021 IEEE International Intelligent Transportation Systems Con- ference (ITSC), 2021, pp. 660–666

work page 2021
[7]

Robust lane change decision for autonomous vehicles in mixed traffic: A safety-aware multi-agent adversarial reinforcement learning approach,

T. Wang, M. Ma, S. Liang, J. Yang, and Y . Wang, “Robust lane change decision for autonomous vehicles in mixed traffic: A safety-aware multi-agent adversarial reinforcement learning approach,” Transporta- tion Research Part C: Emerging Technologies , vol. 172, p. 105005, 2025

work page 2025
[8]

An environment for autonomous driving decision- making,

E. Leurent, “An environment for autonomous driving decision- making,” https://github.com/eleurent/highway-env, 2018

2018

Show all 35 references

[9]

Deeptraffic: Crowdsourced hyperparameter tuning of deep reinforcement learning systems for multi-agent dense traffic navigation,

L. Fridman, J. Terwilliger, and B. Jenik, “Deeptraffic: Crowdsourced hyperparameter tuning of deep reinforcement learning systems for multi-agent dense traffic navigation,” 2019. [Online]. Available: https://arxiv.org/abs/1801.02805

2019 arXiv
[10]

Investigating high-level decision making for automated driving,

A. Capello, L. Forneris, A. Pighetti, F. Bellotti, L. Lazzaroni, M. Cossu, A. De Gloria, and R. Berta, “Investigating high-level decision making for automated driving,” in Applications in Electronics Pervading Industry, Environment and Society, R. Berta and A. De Glo- ria, Eds...

2023
[11]

Designing an interpretability analysis framework for deep reinforcement learning (drl) agents in highway automated driving simulation,

F. Bellotti, L. Lazzaroni, A. Capello, M. Cossu, A. De Gloria, and R. Berta, “Designing an interpretability analysis framework for deep reinforcement learning (drl) agents in highway automated driving simulation,” in Proceedings of SIE 2022 , G. Cocorullo, F. Crupi, and E. Lim...

2022
[12]

A multi-agent reinforcement learning approach for safe and efficient behavior planning of connected autonomous vehicles,

S. Han, S. Zhou, J. Wang, L. Pepin, C. Ding, J. Fu, and F. Miao, “A multi-agent reinforcement learning approach for safe and efficient behavior planning of connected autonomous vehicles,” 2022. [Online]. Available: https://arxiv.org/abs/2003.04371

2022 arXiv
[13]

Safe decision-making for lane-change of autonomous vehicles via human demonstration-aided reinforcement learning,

J. Wu, W. Huang, N. de Boer, Y . Mo, X. He, and C. Lv, “Safe decision-making for lane-change of autonomous vehicles via human demonstration-aided reinforcement learning,” 2022. [Online]. Available: https://arxiv.org/abs/2207.00448

2022 arXiv
[14]

Hierarchical reinforcement learning: A comprehensive survey,

S. Pateria, B. Subagdja, A.-h. Tan, and C. Quek, “Hierarchical reinforcement learning: A comprehensive survey,” ACM Comput. Surv., vol. 54, no. 5, Jun. 2021. [Online]. Available: https: //doi.org/10.1145/3453160

2021 doi
[15]

Trajectory planning for autonomous vehicles using hierarchical reinforcement learning,

K. B. Naveed, Z. Qiao, and J. M. Dolan, “Trajectory planning for autonomous vehicles using hierarchical reinforcement learning,”
[16]

Action and trajectory planning for urban autonomous driving with hierarchical reinforcement learning,

X. Lu, F. X. Fan, and T. Wang, “Action and trajectory planning for urban autonomous driving with hierarchical reinforcement learning,”
[17]

Jointly learnable behavior and trajectory planning for self-driving vehicles,

A. Sadat, M. Ren, A. Pokrovsky, Y .-C. Lin, E. Yumer, and R. Urtasun, “Jointly learnable behavior and trajectory planning for self-driving vehicles,” 2019. [Online]. Available: https://arxiv.org/abs/1910.04586

2019 arXiv
[18]

Combining decision making and trajec- tory planning for lane changing using deep reinforcement learning,

S. Li, C. Wei, and Y . Wang, “Combining decision making and trajec- tory planning for lane changing using deep reinforcement learning,” IEEE Transactions on Intelligent Transportation Systems , vol. 23, no. 9, pp. 16 110–16 136, 2022

2022
[19]

Reinforcement learning,

Wikipedia, “Reinforcement learning,” 2025. [Online]. Available: https://en.wikipedia.org/wiki/Reinforcement learning

2025
[20]

Reinforcement learning: An introduction,

R. Sutton and A. Barto, “Reinforcement learning: An introduction,” IEEE Transactions on Neural Networks , vol. 9, no. 5, pp. 1054–1054, 1998

1998
[21]

S. V . Albrecht, F. Christianos, and L. Sch ¨afer, Multi-Agent Reinforcement Learning: F oundations and Modern Approaches . MIT Press, 2024. [Online]. Available: https://www.marl-book.com

2024
[22]

An efficient centralized multi-agent reinforcement learner for cooperative tasks,

D. Liao, Z. Zhang, T. Song, and M. Liu, “An efficient centralized multi-agent reinforcement learner for cooperative tasks,” IEEE Access, vol. 11, pp. 139 284–139 294, 2023

2023
[23]

Map- based multi-policy reinforcement learning: Enhancing adaptability of robots by deep reinforcement learning,

A. Kume, E. Matsumoto, K. Takahashi, W. Ko, and J. Tan, “Map- based multi-policy reinforcement learning: Enhancing adaptability of robots by deep reinforcement learning,” 2017. [Online]. Available: https://arxiv.org/abs/1710.06117

2017 arXiv
[24]

Multi-policy deep reinforcement learning for multi-objective multiplicity flexible job shop scheduling,

L. Ding, Z. Guan, M. Rauf, and L. Yue, “Multi-policy deep reinforcement learning for multi-objective multiplicity flexible job shop scheduling,” Swarm and Evolutionary Computation , vol. 87, p. 101550, 2024. [Online]. Available: https://www.sciencedirect.com/sc ience/article/p...

2024
[25]

Long short-term memory,

S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997

1997
[26]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,”
[27]

Educational applications of the cyber-physical mobility lab: A summary,

S. Sch ¨afer, J. Xu, D. Kl ¨uner, A. Mokhtarian, P. Scheffel, and B. Alri- faee, “Educational applications of the cyber-physical mobility lab: A summary,” in 2024 European Control Conference (ECC) , 2024, pp. 2666–2671

2024
[28]

Sigmarl: A sample-efficient and gen- eralizable multi-agent reinforcement learning framework for motion planning,

J. Xu, P. Hu, and B. Alrifaee, “Sigmarl: A sample-efficient and gen- eralizable multi-agent reinforcement learning framework for motion planning,” in 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC) , 2024, pp. 768–775

2024
[29]

Available: https://arxiv.org/abs/1706.03762

[Online]. Available: https://arxiv.org/abs/1706.03762

arXiv
[30]

Learning-based control barrier function with provably safe guarantees: Reducing conservatism with heading-aware safety margin,

J. Xu and B. Alrifaee, “Learning-based control barrier function with provably safe guarantees: Reducing conservatism with heading-aware safety margin,” in European Control Conference (ECC), in Press , 2025

2025
[31]

A real-time control barrier function- based safety filter for motion planning with arbitrary road boundary constraints,

J. Xu, C. Che, and B. Alrifaee, “A real-time control barrier function- based safety filter for motion planning with arbitrary road boundary constraints,” in 2025 IEEE 28th International Conference on Intelli- gent Transportation Systems (ITSC), in Press , 2025

2025
[32]

Lanelets: Efficient map repre- sentation for autonomous driving,

P. Bender, J. Ziegler, and C. Stiller, “Lanelets: Efficient map repre- sentation for autonomous driving,” in 2014 IEEE Intelligent V ehicles Symposium Proceedings, 8-11 June 2014, Dearborn, MI, USA . Insti- tute of Electrical and Electronics Engineers (IEEE), 2014, p. 420–425

2014
[35]

High-order control barrier functions: In- sights and a truncated taylor-based formulation,

J. Xu and B. Alrifaee, “High-order control barrier functions: In- sights and a truncated taylor-based formulation,” arXiv preprint arXiv:2503.15014, 2025

2025 arXiv
[2020]

Available: https://arxiv.org/abs/2011.04752

[Online]. Available: https://arxiv.org/abs/2011.04752

2011 arXiv
[2023]

Available: https://arxiv.org/abs/2306.15968

[Online]. Available: https://arxiv.org/abs/2306.15968

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Perception, planning, control, and coordination for autonomous vehicles,

S. D. Pendleton, H. Andersen, X. Du, X. Shen, M. Meghjani, Y . H. Eng, D. Rus, and M. H. Ang, “Perception, planning, control, and coordination for autonomous vehicles,” Machines, vol. 5, no. 1, 2017

work page 2017

[2] [2]

A comprehensive review on safe reinforcement learning for au- tonomous vehicle control in dynamic environments,

R. Inamdar, S. K. Sundarr, D. Khandelwal, V . D. Sahu, and N. Katal, “A comprehensive review on safe reinforcement learning for au- tonomous vehicle control in dynamic environments,” e-Prime - Ad- vances in Electrical Engineering, Electronics and Energy , vol. 10, p. 100810, 2024. (a) Timestep 1106. (b) Timestep 1107. (c) Timestep 1108. Fig. 5: Hesitation...

work page 2024

[3] [3]

Review of decision-making and planning approaches in automated driving,

F. Garrido and P. Resende, “Review of decision-making and planning approaches in automated driving,” IEEE Access, vol. 10, pp. 100 348– 100 366, 2022

work page 2022

[4] [4]

Survey on artificial intelligence for vehicles,

J. Li, H. Cheng, H. Guo, and S. Qiu, “Survey on artificial intelligence for vehicles,” Automotive Innovation, vol. 1, pp. 2–14, 2018

work page 2018

[5] [5]

Decision-making technology for autonomous vehicles: Learning-based methods, applications and future outlook,

Q. Liu, X. Li, S. Yuan, and Z. Li, “Decision-making technology for autonomous vehicles: Learning-based methods, applications and future outlook,” in 2021 IEEE International Intelligent Transportation Systems Conference (ITSC) , 2021, pp. 30–37

work page 2021

[6] [6]

Rule-based decision-making system for autonomous vehicles at intersections with mixed traffic environment,

A. Aksjonov and V . Kyrki, “Rule-based decision-making system for autonomous vehicles at intersections with mixed traffic environment,” in 2021 IEEE International Intelligent Transportation Systems Con- ference (ITSC), 2021, pp. 660–666

work page 2021

[7] [7]

Robust lane change decision for autonomous vehicles in mixed traffic: A safety-aware multi-agent adversarial reinforcement learning approach,

T. Wang, M. Ma, S. Liang, J. Yang, and Y . Wang, “Robust lane change decision for autonomous vehicles in mixed traffic: A safety-aware multi-agent adversarial reinforcement learning approach,” Transporta- tion Research Part C: Emerging Technologies , vol. 172, p. 105005, 2025

work page 2025

[8] [8]

An environment for autonomous driving decision- making,

E. Leurent, “An environment for autonomous driving decision- making,” https://github.com/eleurent/highway-env, 2018

2018