{"id":"343991af-451b-4a77-8eb4-c716a8c25cf4","arxiv_id":"2505.03558","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"In ns-3 simulations, MAPPO with greedy allocation lowers average end-to-end latency and raises latency-success probability versus round-robin for teleoperated driving.","lead":"This paper tests whether multi-agent reinforcement learning can schedule radio resources to lower latency for teleoperated driving. It finds that a centralized learning approach combined with greedy resource allocation performs best in simulation, especially with more vehicles.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim overstates MAPPO+GA benefit: it is worst on the 95th-percentile latency-violation metric that matters for TD reliability (Fig. 4).","rationale":"The reader correctly noted in the strongest_claim that Fig. 4 shows MAPPO+GA has the worst 95th-percentile violation probability, and the reader's rationale mentions the abstract overstates the result. However, the reader's weakest_assumption focuses on the wired backhaul delay rather than this internal metric mismatch. In my view, the tail-latency contradiction is more load-bearing: it does not depend on an external parameter value, but is a direct inconsistency between the paper's unqualified claim and its own reported results. The backhaul sensitivity is a legitimate external-validity concern, but a constant delay added to all schemes would not by itself overturn the average-latency ranking; the tail metric already does. A conditional acceptance is still appropriate because the underlying method and experiments are reproducible and the contribution can be salvaged by qualifying the claim and reporting tail-aware or reliability-based metrics. The proposed concrete test is inexpensive since the data and code exist, and it directly settles whether the headline claim survives under the QoS definition that motivates the paper.","tokens_in":11121,"tokens_out":7951,"duration_ms":82458,"concrete_test":"Using the already-trained MAPPO+GA, MAPPO+PA, and RR policies from the released ns-3 code, recompute per-vehicle 99th-percentile E2E latency and the fraction of vehicles satisfying the 5GAA reliability target P(ell <= tau) >= 0.99, for the N in {5, 8} and (q,c) in {(8,0), (10,10)} configurations. If MAPPO+GA does not remain the best under this reliability-oriented metric, the headline claim must be revised to refer only to average/median latency, not to latency in general.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's stated QoS context (Sec. I) follows 5GAA requirements of 99%-99.999% reliability for teleoperated driving, yet the abstract and Sec. IV-B claim that 'MAPPO, combined with GA, achieves the best results in terms of latency' without qualification. The evidence supports this only for average/median latency and average latency-success probability (Figs. 3 and 5). Fig. 4 directly contradicts an unqualified reading: MAPPO+GA has the worst 95th-percentile latency-violation probability for every N, because GA allocates all resources to the highest-priority UE and starves the rest. Since the tail behavior is decisive for safety-critical TD, the central claim is internally inconsistent with the paper's own results. The conclusion even concedes that 'many severe latency violations are experienced' and that PA promotes fairness. The reward function (2) is threshold-based and penalizes violations only linearly, so optimizing it can favor average success at the expense of tail violations. The claim should be restricted to mean/median latency or the latency-success probability averaged over vehicles, and the abstract should acknowledge the worst-case trade-off.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes multi-agent reinforcement learning (MARL) schedulers for teleoperated driving (TD) networks, in which each vehicle/agent asks the RAN for a priority level and the gNB allocates OFDM symbols either proportionally (PA) or greedily (GA). Two PPO-based training variants are compared: independent PPO (IPPO) and multi-agent PPO (MAPPO). The evaluation, carried out in ns-3 with mmWave traces from GEMV2 and SUMO mobility, compares MAPPO+PA, MAPPO+GA, IPPO variants, and a round-robin (RR) baseline as a function of the number of vehicles, the Draco compression configuration, and the latency threshold. The central claim, stated in the abstract and repeated in Sec. IV-B, is that MAPPO combined with GA achieves the best latency performance, especially as the number of vehicles grows, and that it maximizes the number of vehicles satisfying the latency constraint.","tokens_in":11303,"tokens_out":4265,"duration_ms":47904,"significance":"If understood as a claim about average/median latency and average latency-success probability, the paper makes a useful, reproducible empirical contribution: it releases source code, uses open simulators with realistic channel and mobility traces, and provides a concrete MARL formulation of RAN-level scheduling for TD. The comparison between IPPO and MAPPO under two allocation rules is also informative. However, the headline claim as written is broader than the evidence: the paper's own Fig. 4 shows that MAPPO+GA has the worst tail behavior (95th-percentile latency-violation probability) for every number of vehicles, and the conclusion explicitly concedes that many severe latency violations occur. Given that the paper motivates TD with 99%-99.999% reliability requirements, tail behavior is not a secondary concern but a decisive one. The contribution is therefore solid but needs a substantially more careful statement of what exactly is optimized and what the trade-offs are.","major_comments":[{"comment":"The unqualified claim that \"MAPPO, combined with GA, achieves the best results in terms of latency\" is contradicted by the paper's own Fig. 4, which shows that MAPPO+GA has the highest 95th-percentile latency-violation probability for every N (3, 5, and 8). Since Sec. I motivates the work with 5GAA reliability requirements of 99% to 99.999%, the tail metric is the most safety-relevant one, not a minor caveat. Please restrict the central claim to average/median latency and to the average latency-success probability, and explicitly state in the abstract and in Sec. IV-B-b that MAPPO+GA trades fairness and worst-case violations for average performance.","section":"Abstract; Sec. IV-B-b"},{"comment":"All comparative conclusions are based on point estimates, but no confidence intervals, error bars, or statistical significance tests are reported despite 250 independent simulation runs per setting. For example, Sec. IV-B-b states that at N=5 MAPPO+GA improves the latency-success probability by about 35% over RR, yet no variance information is given, and Fig. 4 also reports point estimates without uncertainty. Without such information, \"best\" is not statistically established, especially because the reported differences in some regimes are small (e.g., the reward values in Fig. 1). Please add confidence intervals or significance tests for the key metrics, or at minimum error bars on the figures.","section":"Sec. IV-A-e; Figs. 3-5"},{"comment":"The reward function gives a constant +1 whenever latency is at or below the threshold and a linear penalty proportional to the violation size otherwise. This reward rewards reducing the average violation count but does not distinguish a moderate violation from a severe tail violation, and it provides no gradient toward avoiding extreme latency. This objective mismatch is a likely mechanism behind the Fig. 4 result, where the policy that maximizes average reward is also the one with the worst 95th-percentile violation probability. Please discuss this trade-off explicitly and consider either a tail-sensitive reward term or a per-UE penalty structure that prevents systematic starvation of lower-priority vehicles.","section":"Sec. III-A-c, Eq. (2)"},{"comment":"The wired backhaul is modeled with a fixed propagation delay of 10 ms and a rate of 100 Gbps. This is a large, configuration-independent component of the 15-35 ms end-to-end latency budget, and although it may not change the relative ordering of schedulers, it can determine whether a configuration meets the threshold at all (e.g., the claim that MAPPO+GA is the only method keeping median latency below the threshold for (10,10) depends on the absolute latency budget). Please justify the 10 ms value, clarify whether it is one-way or round-trip, and ideally provide a sensitivity sweep over backhaul delay and rate to show that the qualitative conclusions are robust.","section":"Sec. IV-A-a"}],"minor_comments":[{"comment":"The phrase \"we prove via ns-3 simulations\" is too strong for simulation-based evidence; \"demonstrate\" or \"show\" would be more appropriate.","section":"Abstract"},{"comment":"The number of priority levels K=3 is chosen empirically, but no sensitivity analysis is given. Since the greedy allocation rule starves lower-priority UEs by construction, the sensitivity of the tail results to K should at least be discussed.","section":"Sec. IV-A-c"},{"comment":"The end-to-end latency definition says it is measured from packet generation at the transmitter to reception, but it is not stated explicitly whether this includes the wired backhaul delay and queueing at the remote host or only the radio access segment. Please make the decomposition of the E2E latency explicit.","section":"Sec. IV-A-e"},{"comment":"The top subplot is described in the text as a boxplot, but the figure caption does not explain the boxplot elements (median, quartiles, whiskers). Please add a sentence to the caption.","section":"Fig. 3"},{"comment":"The sentence \"the median latency is always lower than tau with the current settings\" is stated before reporting the boxplot; please clarify that this holds for the evaluated configurations and that the mean can still exceed tau due to tail violations.","section":"Sec. IV-B-b"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-scoped and the reproducibility effort (open source, ns-3, realistic traces) is a genuine strength. The main issue for me is not the methodology but the framing: the headline claim is contradicted by the paper's own tail metric, and the conclusion already concedes the fairness problem. This is fixable with a clearer claim, a discussion of the objective mismatch, and ideally uncertainty quantification. The paper also cites the authors' prior work extensively, but those references are directly relevant to the PQoS/RAN-AI setup, so I do not see this as a concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Have you seen this one? The abstract says MAPPO+GA achieves the best latency and 'proves' it via simulation, but their own Figure 4 shows MAPPO+GA has the worst 95th-percentile latency-violation probability for every N. The paper is still worth a look: it applies MARL to RAN-level OFDM symbol scheduling for teleoperated driving, a sensible extension of the authors' earlier compression-level RL work, and they ship their ns-3 code.\n\nThe new part is the scheduling-level control: IPPO vs MAPPO under proportional vs greedy allocation, with RR as baseline. The simulation setup is reasonably careful: mmWave traces from GEMV2, SUMO mobility, 250 independent runs per configuration. The reward function is standard and not circular.\n\nThe soft spot is not minor. The abstract and Section IV-B claim MAPPO+GA gives the best latency, but the tail metric shows the opposite, and the conclusion admits 'many severe latency violations' with this scheme. Given the paper's own 99%-99.999% reliability requirement, the claim should be restricted to mean/median latency and average latency-success probability, and the trade-off should be in the abstract.\n\nTwo smaller issues: the wired backhaul is 'ideal' with a 10 ms propagation delay, which may dominate the E2E budget and could change the scheduler ranking if that delay were smaller; and there are no comparisons with other learning-based schedulers or confidence intervals, despite 250 runs.\n\nMy take: this is an incremental but real contribution, honestly presented except for the overclaim. It deserves peer review, not desk rejection. The fix is straightforward: qualify the claim, add a backhaul sensitivity analysis, and report variance. That would make it a solid conference paper and a plausible journal candidate.","headline":"MAPPO+GA is the best on average latency but worst on the 95th-percentile tail in the paper's own Figure 4, so the abstract overclaims; the underlying study is solid enough to save.","tokens_in":11865,"tokens_out":3350,"would_cite":false,"duration_ms":32429,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Radio-resource scheduling trained by multi-agent reinforcement learning can keep teleoperated driving inside its end-to-end latency budget without compressing the sensor data.","keywords":["teleoperated driving","predictive quality of service","multi-agent reinforcement learning","proximal policy optimization","radio resource scheduling","end-to-end latency","5G NR","mmWave"],"falsifier":"Run the same MAPPO-with-greedy-allocation versus round-robin comparison across backhaul delays of 1, 5, 10, and 20 ms in the simulator, or on a hardware-in-the-loop teleoperation testbed with a measured 5G data path; if the approximately 35% latency-success advantage at five vehicles and aggressive compression disappears when the backhaul delay drops to 1 ms, the paper's claim that MARL scheduling is the decisive lever for end-to-end latency would be falsified.","tokens_in":10892,"feed_emoji":"🚗","tokens_out":9456,"duration_ms":84030,"temperature":0.7,"pith_summary":"The paper tries to show that the radio scheduler, not data compression, is the right lever for keeping teleoperated driving within its end-to-end latency budget. It trains multiple agents with Proximal Policy Optimization to assign radio resources (OFDM symbols per 5G slot) to vehicles according to priority levels, and compares independent training (IPPO) with centralized training (MAPPO) under both proportional and greedy allocation. The central claim is that MAPPO combined with greedy allocation gives the lowest average latency and the highest probability of meeting the latency threshold, with the advantage growing as the number of vehicles and the data rate increase. A reader should care because this approach preserves the quality of the LiDAR data being transmitted while still reacting to congested or degraded network conditions.","feed_headline":"AI radio scheduler beats round-robin for teleoperated driving","feed_subtitle":"Centralized multi-agent PPO plus greedy allocation hits latency targets that round-robin misses, at higher vehicle counts.","key_machinery":"The load-bearing mechanism is the MARL scheduler operating inside the radio access network. Each vehicle is an agent whose observation bundles average SINR, uplink buffer size, required OFDM symbols, average modulation and coding index, and recent end-to-end latency; its action is a priority level k in {1,...,K}; and the reward is +1 when end-to-end latency is at or below the threshold and a penalty proportional to the violation otherwise. The scheduling strategies convert priorities into OFDM-symbol allocations: proportional allocation weights each vehicle's share by its priority and gives leftover symbols to the highest-priority vehicles, while greedy allocation gives all symbols to the highest-priority vehicle before serving the next. MAPPO is the centralized-training, decentralized-execution variant in which one actor-critic pair is trained on data from all agents, and the paper argues that this shared model coordinates agents better than IPPO, which trains each agent on its own observations.","core_discovery":"On the paper's own terms, the discovery is that a multi-agent reinforcement-learning scheduler trained with centralized parameter sharing (MAPPO) and allocating all available OFDM symbols to the most latency-constrained vehicle (greedy allocation) outperforms both a round-robin baseline and decentralized training for teleoperated driving over millimeter-wave links. In simulations with five vehicles and the most aggressive compression setting, the probability that end-to-end latency stays under threshold is about 35% higher than with round-robin; under the least compressed setting, only MAPPO with greedy allocation keeps the median latency below the threshold, satisfying it more than 70% of the time versus about 50% for proportional allocation and 42% for round-robin. The paper also reports that this same greedy variant has the worst 95th-percentile latency-violation probability, because it starves lower-priority vehicles; the authors present the result as the best average-latency performance at the cost of fairness.","pith_inferences":["The fixed 10 ms backhaul propagation delay in the simulation is a large share of the 25 ms latency budget, so I expect the relative scheduler ranking to compress if that delay is reduced; changing only the backhaul delay in the simulator would test this directly.","A reward function that penalizes tail latency more heavily, such as the 95th percentile rather than the average, would likely change the ranking between greedy and proportional allocation, given that the paper's own tail-violation results favor proportional allocation.","A natural extension is to let the number of priority levels scale with the number of vehicles or to use continuous actions, which could reduce the starvation that greedy allocation causes while keeping its average-latency benefit.","The RAN-level scheduler could be paired with compression as a fallback: use MARL scheduling first and activate data compression only when the predicted latency still violates the threshold, bounding the data-quality loss while keeping latency under control."],"forward_implications":["Network operators could improve teleoperated-driving latency without modifying the sensor data, since scheduling alone can lift the share of vehicles meeting a 25 ms end-to-end budget.","MAPPO with greedy allocation is the recommended configuration when congestion comes from many vehicles or high-rate LiDAR traffic, while under light load it offers little or no gain over round-robin.","The average-latency win of greedy allocation comes with a fairness cost: it produces the worst 95th-percentile latency violations, so delay-sensitive applications that care about worst cases should prefer proportional allocation or a hybrid strategy.","As the number of vehicles grows, all schedulers converge in average latency because the system becomes resource-limited, but the MARL schedulers still keep a higher latency-success probability than round-robin.","Because the action space is a small set of priority levels rather than per-symbol assignments, the approach fits existing 5G NR frame structures and remains scalable as the network grows."],"supporting_citations":[{"why":"It defines the teleoperated-driving latency and reliability requirements that set the evaluation thresholds.","marker":"[3]"},{"why":"It frames Predictive Quality of Service and the RAN-level optimization concept this paper builds on.","marker":"[5]"},{"why":"It establishes the teleoperated-driving scenario with LiDAR point-cloud traffic and compression configurations that this paper reuses.","marker":"[6]"},{"why":"It supplies the open-source 5G NR millimeter-wave protocol-stack simulator used for all network-level evaluations.","marker":"[12]"},{"why":"It generates the urban millimeter-wave channel traces that drive the simulated propagation conditions.","marker":"[14]"},{"why":"It provides the RAN-AI entity and simulator integration that let the reinforcement-learning agents control the scheduler.","marker":"[16]"},{"why":"It supplies the Proximal Policy Optimization algorithm whose multi-agent variants are trained in this paper.","marker":"[20]"},{"why":"It defines Independent PPO, the decentralized training baseline that this paper compares against centralized training.","marker":"[23]"},{"why":"It defines Multi-Agent PPO, the centralized-training method at the core of the best-performing scheduler.","marker":"[24]"}],"fun_headline_variants":["Centralized multi-agent RL scheduler cuts teleoperated driving latency","MAPPO with greedy allocation beats round-robin for TD latency","AI scheduler for teleoperated driving: centralized training wins","Greedy multi-agent RL minimizes latency in teleoperated driving","Centralized PPO radio scheduler outperforms round-robin in TD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulated environment, especially an ideal wired backhaul with a fixed 10 ms propagation delay and high-fidelity millimeter-wave channel traces from an urban mobility scenario, is representative enough that the relative ordering of schedulers carries over to real teleoperated driving; if the real backhaul delay is much smaller than 10 ms, the radio scheduler's share of the end-to-end latency budget changes and the ranking could shift.","fun_headline_variants_meta":{"raw":{"variants":["Centralized multi-agent RL scheduler cuts teleoperated driving latency","MAPPO with greedy allocation beats round-robin for TD latency","AI scheduler for teleoperated driving: centralized training wins","Greedy multi-agent RL minimizes latency in teleoperated driving","Centralized PPO radio scheduler outperforms round-robin in TD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1480,"prompt_tokens":990,"completion_tokens":490,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":403}},"tokens_in":606,"tokens_out":490,"duration_ms":4276,"temperature":1.0,"reasoning_tokens":403,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:48:08.948541+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same MAPPO-with-greedy-allocation versus round-robin comparison across backhaul delays of 1, 5, 10, and 20 ms in the simulator, or on a hardware-in-the-loop teleoperation testbed with a measured 5G data path; if the approximately 35% latency-success advantage at five vehicles and aggressive compression disappears when the backhaul delay drops to 1 ms, the paper's claim that MARL scheduling is the decisive lever for end-to-end latency would be falsified.","supporting_citations":[{"cited_title":"C-V2X Use Cases V olume II: Examples and Service Level Requirements,","cited_arxiv_id":null,"evidence_quote":"It defines the teleoperated-driving latency and reliability requirements that set the evaluation thresholds."},{"cited_title":"Predictive Quality of Service (PQoS): The Next Frontier for Fully Autonomous Systems,","cited_arxiv_id":null,"evidence_quote":"It frames Predictive Quality of Service and the RAN-level optimization concept this paper builds on."},{"cited_title":"A Reinforcement Learning Framework for PQoS in a Teleoperated Driving Scenario,","cited_arxiv_id":null,"evidence_quote":"It establishes the teleoperated-driving scenario with LiDAR point-cloud traffic and compression configurations that this paper reuses."},{"cited_title":"End-to-End Simulation of 5G mmWave Networks,","cited_arxiv_id":null,"evidence_quote":"It supplies the open-source 5G NR millimeter-wave protocol-stack simulator used for all network-level evaluations."},{"cited_title":"Geometry-based vehicle-to- vehicle channel modeling for large-scale simulation,","cited_arxiv_id":null,"evidence_quote":"It generates the urban millimeter-wave channel traces that drive the simulated propagation conditions."},{"cited_title":"Artificial Intelligence in Vehicular Wireless Networks: A Case Study Using ns-3,","cited_arxiv_id":null,"evidence_quote":"It provides the RAN-AI entity and simulator integration that let the reinforcement-learning agents control the scheduler."},{"cited_title":"The surprising effectiveness of PPO in cooperative multi-agent games,","cited_arxiv_id":null,"evidence_quote":"It defines Multi-Agent PPO, the centralized-training method at the core of the best-performing scheduler."}],"review_version":1}