{"id":"f3af2bb0-2379-402f-b6d2-c92e29e15da9","arxiv_id":"2506.00982","paper_version":3,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"RSR-RSMARL is a robust safe MARL framework with V2V communication and CBF safety shields that supports zero-shot sim-to-real transfer and improves coordination on 1/10-scale vehicle hardware.","lead":"This paper introduces RSR-RSMARL, a framework combining robust multi-agent reinforcement learning with communication and control barrier functions to enable safe policy transfer from simulation to real hardware for autonomous vehicles. Smart generalists might read it to see practical steps toward deploying coordinated AI driving systems beyond simulation.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Zero-shot hardware transfer rests on unquantified sim-to-real gap closure via state/action representations alone","rationale":"The reader's weakest assumption directly isolates the transfer mechanism; the full text does not add quantitative evidence that closes this gap, so the concern remains load-bearing for the hardware demonstration claim. This moves the verdict from UNVERDICTED to CONDITIONAL pending the proposed distribution check.","tokens_in":1776,"tokens_out":338,"duration_ms":28021,"concrete_test":"Re-run the hardware trials (Section 5) while logging the exact state vectors and CBF intervention rates for the same policy weights used in simulation; compute Wasserstein distance or KL divergence between sim and hardware state distributions for the shared V2V features. If divergence exceeds the robustness training bound reported in Section 3.3 and hardware safety metrics drop below simulation by >15%, the representation-based transfer claim does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that the chosen state (including shared V2V info) and action representations, plus robust MARL training, close the sim-to-real gap sufficiently for direct deployment on 1/10-scale vehicles. Section 4 and the RSR adaptation description do not report explicit gap metrics (e.g., state distribution divergence, actuator delay mismatch, or sensor noise statistics between sim and hardware). Without these, it is unclear whether observed hardware success stems from the representations or from unstated tuning, environment simplification, or the CBF shield overriding policy actions. If the gap is larger than the robustness margin, the zero-shot guarantee fails even if simulation results look strong.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes RSR-RSMARL, a novel Robust and Safe Multi-Agent Reinforcement Learning framework with V2V communication for autonomous vehicles. It enables Real-Sim-Real (RSR) policy adaptation by designing state (including shared information) and action representations that account for real-system complexities, training via a robust MARL algorithm for zero-shot hardware transfer, and adding a Control Barrier Function (CBF) safety shield per agent. The central claim is that this yields enhanced driving safety and coordination, supported by both simulation results and hardware experiments on 1/10th-scale vehicles across multiple configurations.","tokens_in":1901,"tokens_out":501,"duration_ms":19011,"significance":"If the hardware results hold with quantitative support, the work would be significant for multi-agent autonomy: it directly tackles sim-to-real transfer, communication design, and safety in a single modular architecture. The combination of representation choices, robust training, and CBF shielding offers a concrete path toward deployable MARL policies on physical vehicles, which remains rare in the literature.","major_comments":[{"comment":"Abstract and Section 4: The manuscript asserts that 'Experimental results on 1/10th-scale autonomous vehicles with V2V communication demonstrate the ability of RSR-RSMARL framework to enhance driving safety and coordination,' yet supplies no quantitative metrics (success rates, collision counts, trajectory error, or statistical significance), no baselines, and no error analysis or training hyperparameters. This absence directly undermines the central empirical claim of effective zero-shot transfer.","section":"Abstract and Section 4"},{"comment":"Section 4 and RSR adaptation description: No explicit quantification of the sim-to-real gap is provided (e.g., Wasserstein distance between state distributions, actuator latency mismatch, or sensor noise statistics). Without these measurements it is impossible to determine whether any observed hardware performance arises from the chosen state/action representations or from unstated environment simplifications or CBF intervention. This is load-bearing for the zero-shot guarantee.","section":"Section 4 and RSR adaptation description"}],"minor_comments":[{"comment":"The expansion of the acronym RSR-RSMARL is not stated on first use; adding '(Robust and Safe Real-Sim-Real Multi-Agent Reinforcement Learning)' would improve readability.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"Thank you for the opportunity to respond to the referee's report. We value the constructive criticism regarding the presentation of our hardware experiments and the quantification of the sim-to-real gap. We believe these points can be addressed through targeted revisions and additional analysis, which we outline below.","responses":[{"response":"We acknowledge the validity of this observation. The current manuscript emphasizes qualitative demonstrations and figures in Section 4 to illustrate the hardware performance. In the revision, we will incorporate quantitative metrics such as success rates, number of collisions, trajectory errors with standard deviations, and p-values for statistical significance. Baselines including non-communicative MARL and MARL without CBF will be added, along with a table summarizing hyperparameters and error analysis. This will provide the necessary quantitative support for the zero-shot transfer claims.","revision_made":"yes","referee_comment":"[Abstract and Section 4] Abstract and Section 4: The manuscript asserts that 'Experimental results on 1/10th-scale autonomous vehicles with V2V communication demonstrate the ability of RSR-RSMARL framework to enhance driving safety and coordination,' yet supplies no quantitative metrics (success rates, collision counts, trajectory error, or statistical significance), no baselines, and no error analysis or training hyperparameters. This absence directly undermines the central empirical claim of effective zero-shot transfer."},{"response":"We agree that providing explicit measures of the sim-to-real gap would enhance the rigor of our claims. We will revise Section 4 to include an analysis of the sim-to-real discrepancies, such as statistical comparisons of state distributions (including Wasserstein distance where applicable), measured actuator latencies, and sensor noise levels from the hardware setup. We will also clarify how the designed state and action representations mitigate these gaps and evaluate the contribution of the CBF safety shield through ablation studies. These additions will better justify the zero-shot transfer performance.","revision_made":"yes","referee_comment":"[Section 4 and RSR adaptation description] Section 4 and RSR adaptation description: No explicit quantification of the sim-to-real gap is provided (e.g., Wasserstein distance between state distributions, actuator latency mismatch, or sensor noise statistics). Without these measurements it is impossible to determine whether any observed hardware performance arises from the chosen state/action representations or from unstated environment simplifications or CBF intervention. This is load-bearing for the zero-shot guarantee."}],"tokens_in":1503,"tokens_out":510,"duration_ms":43316,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The core takeaway is that this work shows a complete pipeline—robust multi-agent RL training, shared state representations via V2V, real-sim-real adaptation, and per-agent CBF safety modules—running on 1/10th-scale cars with actual communication hardware. That combination is more than most MARL papers deliver, and the hardware step itself is useful to see even at small scale. They get credit for treating communication as part of the state rather than an afterthought and for keeping the safety layer modular so it does not interfere with the learned policy during training. The architecture description is clear enough that someone could replicate the high-level structure. The abstract also notes success across multiple configurations, which suggests the framework is not tuned to one narrow case. The main weakness is the missing numbers. No success rates, collision counts, baseline comparisons, or direct measurements of the sim-to-real gap appear in the provided material. Without those, it is difficult to judge whether the state and action representations actually closed the transfer gap or whether the CBF shield simply overrode risky actions on hardware. The stress-test concern about unquantified discrepancies in state distributions, delays, and noise holds up here; if those details are in the full paper they are not highlighted, which weakens the zero-shot claim. This paper is aimed at researchers working on multi-agent autonomy who already know the MARL and CBF literature and want to see an end-to-end hardware attempt. A reader in that group would find the integration and the small-robot demonstration worth reading, even if they would still need to run their own experiments to trust the transfer results. I would send it to peer review. The hardware component and the modular safety design are concrete enough to justify referee time, provided the authors supply the quantitative metrics and gap analysis that are currently absent.","headline":"The paper integrates robust MARL with V2V communication and CBF safety shields for a hardware demo on small-scale vehicles, but the sim-to-real transfer claims rest on thin evidence.","tokens_in":2401,"tokens_out":440,"would_cite":false,"duration_ms":22830,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[{"relation":"unclear","rs_module":"IndisputableMonolith/Foundation/Cost/FunctionalEquation.lean","rs_theorem":"washburn_uniqueness_aczel","paper_passage":"RSR-RSMARL leverages state (includes shared state information among agents) and action representations considering real system complexities for MARL formulation. The MARL policy is trained with robust MARL algorithm to enable zero-shot transfer to hardware considering the sim-to-real gap. A safety shield module using Control Barrier Functions (CBFs) provides safety guarantee for each individual agent."},{"relation":"unclear","rs_module":"IndisputableMonolith/Foundation/AlexanderDuality.lean","rs_theorem":"alexander_duality_circle_linking","paper_passage":"We adopt the kinematic bicycle model ... The CBF is the additional safety constraint ... min u ½∥u−uref∥² s.t. ∂h/∂t + Lf h + Lg h u ≥ −γh"}],"headline":"Practical MARL sim-to-real pipeline with CBF shields has no structural overlap with RS forcing chain","alignment":"orthogonal","rationale":"Paper centers on CTDE MARL (MAPPO variant), delayed-state robustness, V2V shared observations, and CBF-QP safety filters for 1/10-scale AV hardware transfer. None of these components invoke or parallel RS primitives (J-cost functional equation, φ-ladder, 8-tick periodicity, distinction-to-spacetime forcing). Domain is applied robotics; RS has no theorems about reinforcement learning or control-barrier safety shields.","tokens_in":50881,"confidence":"high","tokens_out":369,"duration_ms":13864,"cache_read_input_tokens":32896,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A MARL framework trains driving policies in simulation and transfers them directly to physical vehicles while adding safety shields.","keywords":["multi-agent reinforcement learning","autonomous vehicles","sim-to-real transfer","vehicle-to-vehicle communication","control barrier functions","robust learning","safety shields","hardware experiments"],"falsifier":"Deploy the simulator-trained policies on the 1/10th-scale vehicles without any fine-tuning and observe whether safety or coordination breaks down in the presence of communication delays, model uncertainties, or dynamic obstacles.","tokens_in":2691,"feed_emoji":"🚗","tokens_out":761,"duration_ms":25539,"temperature":0.7,"pith_summary":"The paper establishes that multi-agent reinforcement learning policies for autonomous vehicles can be formulated with state and action representations that explicitly account for physical system complexities, trained robustly in simulation, and then transferred zero-shot to hardware. It incorporates vehicle-to-vehicle communication for shared information and uses Control Barrier Functions as modular safety shields to enforce guarantees during both training and deployment. A sympathetic reader would care because this addresses the persistent sim-to-real gap and safety concerns that have limited learning-based methods in real multi-robot systems, potentially allowing safer coordinated driving without extensive real-world retraining.","feed_headline":"Framework transfers safe multi-agent driving policies from sim to hardware","feed_subtitle":"State representations, robust training, and barrier-function shields let policies coordinate vehicles on 1/10-scale hardware without retrain","key_machinery":"The RSR-RSMARL framework, which combines robust MARL training, state and action representations that include shared V2V information and real-system details, Real-Sim-Real adaptation, and modular Control Barrier Function safety shields to support zero-shot hardware transfer.","core_discovery":"RSR-RSMARL is a Robust and Safe MARL framework that supports Real-Sim-Real policy adaptation for multi-agent systems with communication among agents. It leverages state representations that include shared information among agents and action representations that consider real system complexities. The policy is trained with a robust MARL algorithm to enable zero-shot transfer to hardware despite the sim-to-real gap. A safety shield module using Control Barrier Functions provides safety guarantees for each individual agent. Experiments on 1/10th-scale autonomous vehicles with V2V communication show that the framework enhances driving safety and coordination across multiple configurations.","pith_inferences":["The approach might scale to full-size vehicles if the state representations are adjusted for higher speeds and longer communication ranges.","Similar combinations of robust training and barrier-function shields could apply to other multi-agent domains such as drone coordination or warehouse robots.","If communication is intermittent, the framework's reliance on shared states would need explicit robustness extensions that the current experiments do not test.","The method could be combined with online adaptation modules to handle larger distribution shifts not seen in the 1/10-scale tests."],"forward_implications":["Multi-agent vehicle teams can maintain individual safety guarantees while using shared communication to improve overall coordination.","Zero-shot transfer from simulation becomes feasible for MARL policies when representations are designed around physical complexities rather than idealized models.","Safety shields based on Control Barrier Functions can be added modularly without retraining the core policy for hardware use.","The same framework supports testing across varied team sizes and scenarios once the representations and training are fixed."],"fun_headline_variants":["MARL transfers safely from sim to 1/10-scale vehicle hardware","Shared states and safety shields advance MARL to hardware vehicles","Robust MARL with V2V communication coordinates real autonomous fleets","Control barrier functions protect sim-to-hardware multi-agent driving"],"cache_read_input_tokens":64,"weakest_assumption_plain":"State and action representations that capture real system complexities, together with robust training, are enough to overcome sim-to-real discrepancies and model uncertainties so that the policies work directly on physical hardware.","fun_headline_variants_meta":{"raw":{"variants":["MARL transfers safely from sim to 1/10-scale vehicle hardware","Shared states and safety shields advance MARL to hardware vehicles","Robust MARL with V2V communication coordinates real autonomous fleets","Control barrier functions protect sim-to-hardware multi-agent driving"]},"model":"grok-4.3","cost_usd":0.009032,"raw_usage":{"total_tokens":4025,"prompt_tokens":771,"num_sources_used":0,"completion_tokens":69,"cost_in_usd_ticks":90315500,"prompt_tokens_details":{"text_tokens":771,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3185,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":771,"tokens_out":69,"duration_ms":44666,"temperature":1.0,"reasoning_tokens":3185,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-19T11:15:46.444624+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Deploy the simulator-trained policies on the 1/10th-scale vehicles without any fine-tuning and observe whether safety or coordination breaks down in the presence of communication delays, model uncertainties, or dynamic obstacles.","supporting_citations":[],"review_version":1}