{"id":"dc1af923-1336-4af2-97f7-0eee85391a7e","arxiv_id":"2509.23960","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MAD-PINN learns a decentralized safety-optimal policy from a physics-informed HJB value function and reports near-collision-free multi-agent navigation up to 256 agents in simulation.","lead":"MAD-PINN trains one neural network on small groups of robots, then lets each robot in a large swarm use that learned value function with a few safety-critical neighbors. In simulated drone navigation, it reports better safety and cost trade-offs than MARL, safety-filter, and MPC baselines, scaling to 256 agents.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Decentralized execution extracts individual controls from a jointly-optimal local value function, but no argument shows the composed trajectories satisfy the local safety condition.","rationale":"The reader's weakest assumption—that interactions beyond the selected n neighbors are irrelevant—is a real and visible risk, supported by the degradation in safe-scenario rates as agent count grows. I partially agree: that truncation is one way the local model can fail to describe the global system. However, the more fundamental unsupported step is that a local joint-optimal value function can be turned into an individual control law by naive extraction. Even with an exact value function and with all agents inside the observation radius, the composed decentralized controls are not the joint-optimal controls for any cluster, so the local safety certificate does not apply to the actual trajectory. The paper asserts global consistency from homogeneity alone, which is not justified. The proposed audit directly checks whether the controls that agent i plans for its neighbors match what those neighbors actually do. If they match, this concern does not land and the reader's truncation concern remains the main caveat; if they do not, the central scalability claim rests on an unsupported equilibrium assumption. This reinforces, rather than overturns, the reader's conditional verdict: the empirical claims may hold in the tested simulator, but the formal 'globally consistent solution' language and any safety-guarantee interpretation should be softened until the consistency mechanism is demonstrated or removed.","tokens_in":12279,"tokens_out":12234,"duration_ms":144996,"concrete_test":"Conduct a consistency audit on the existing 8- and 16-agent rollouts. At each receding-horizon replanning step, for every agent i and each selected neighbor j, record the control \\hat u_j^i that agent i's Hamiltonian minimizer prescribes for j, and compare it with the control u_j^j that j actually executes from j's own minimizer. Report the mean and maximum norm of \\hat u_j^i - u_j^j over all i, j, steps, and episodes. Then re-simulate each local cluster using the actually executed controls and check whether \\hat V_\\theta(t,o,z^*) \\le 0 is preserved along the realized cluster trajectory. If prescribed and executed controls diverge materially, or the local value condition is violated on realized trajectories, the 'globally consistent solution' claim in Section II-A fails for decentralized execution, independent of PINN error or non-neighbor truncation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's bridge from centralized training to decentralized execution is missing. Section II-A defines Vd in (4) as a centralized SC-OCP for the local cluster: u_d is the joint control for all agents in the observation, and the epigraph guarantee \\hat V_d \\le 0 (Eqs. 5-6) certifies that the cluster trajectory under the joint minimizer stays in g_d \\le 0. Section III-C then lets each agent i compute \\pi_i as (the i-component of) the joint minimizer over u_d for its own observation. But agents j,k in i's observation are not executing i's minimizer; they execute controls from their own local Hamiltonians. The realized global trajectory is therefore not the joint-optimal trajectory for any single cluster, and the local safety certificate \\hat V_d \\le 0 does not apply to the actual multi-agent state. Homogeneity alone does not create consistency; the paper gives no fixed-point, equilibrium, or bounded-mismatch argument. This failure mode is independent of PINN approximation error and of truncation to n neighbors: even if every other agent were inside the observation radius, the same inconsistency remains. The empirical drop from 100% safe scenarios at N=3 (where all agents share one cluster) to 86% at N=16 and 85% at N=256 is consistent with, but does not isolate, this issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MAD-PINN, a decentralized physics-informed machine learning framework for multi-agent state-constrained optimal control (SC-OCP). The method trains, on a small 3-agent system, an epigraph-based auxiliary value function satisfying an HJB PDE, using a boundary-aware PINN. For deployment, each agent selects a fixed number of safety-critical neighbors via a DeepReach pairwise value function and executes a receding-horizon policy derived from the shared local value function. Experiments on multi-agent drone navigation compare against DEF-MARL, SafeMARL, and MPPI, reporting superior cumulative-cost/safety trade-offs, high safety rates up to 256 agents, and an ablation supporting the reachability-based neighbor selection. The central claim is that a single locally trained value function, deployed per-agent with local observations, yields near-collision-free, low-cost navigation for arbitrarily many homogeneous agents.","tokens_in":12618,"tokens_out":2652,"duration_ms":24143,"significance":"If the safety guarantee were rigorously established, the paper would make a useful contribution: it demonstrates that an SC-OCP value function trained on a small cluster can be deployed in a decentralized way with good empirical performance, and the receding-horizon neighbor selection is a sensible engineering idea. The experimental comparison is thorough, with standard errors across seeds, and the ablation in Table III directly supports the value of the proposed neighbor-selection mechanism. The boundary-aware network structure that exactly encodes the terminal condition is also a clean design choice. However, the paper's load-bearing claim of 'strict safety guarantees' and 'globally consistent' solutions is not supported by the presented theory. The missing arguments are not cosmetic: they concern whether the decentralized execution actually solves or approximates the intended safe-optimal control problem, and whether the learned PINN approximation carries any certificate. These gaps currently put the formal framing of the paper ahead of its evidence.","major_comments":[{"comment":"The bridge from centralized local training to decentralized execution is missing. In Eq. (4), the value function V_d is defined for a single local cluster, and the certificate in Eqs. (5)-(6) applies to the joint control u_d of that cluster. In Section III-C, each agent i extracts the i-component of the joint minimizer for its own observation. But other agents are not executing i's minimizer; they execute controls from their own local Hamiltonians. The realized multi-agent trajectory is therefore not the joint-optimal trajectory for any single cluster, so the local certificate \\hat V_d <= 0 does not apply to the actual system. The paper provides no fixed-point, equilibrium, or bounded-mismatch argument to close this gap. This is independent of PINN approximation error and of truncation to n neighbors. The empirical degradation from 100% safe scenarios at N=3 to 86% at N=16 (Table I) and","section":"Section III-C and Section II-A"},{"comment":"The formal safety claim relies on the auxiliary value function being the unique viscosity solution of the HJB PDE (8), which is only true under assumptions A1-A4 from [23]. The paper neither states these assumptions nor verifies that the double-integrator drone system satisfies them. Moreover, the learned \\hat V_theta is a PINN approximation minimizing the PDE residual (10) at sampled collocation points; no convergence, verification, or worst-case error bound is provided. The conclusion section acknowledges this by listing 'quantify the approximation error of the auxiliary value function' as future work. Consequently, the abstract's claim of 'strict safety guarantees' is not established. At minimum, the paper should either prove or explicitly cite a verification result for \\hat V_theta, or reframe the contribution as an empirical heuristic without formal guarantees.","section":"Section II-B, Eq. (8), and Section III-A"},{"comment":"The assumption that fixed-size local observations of n neighbors fully determine each agent's safe-optimal decision is asserted but not analyzed. Section II-A claims this yields a 'globally consistent solution' independent of N, but no theorem or bound quantifies the effect of ignoring non-neighbor agents. Receding-horizon re-selection of neighbors does not eliminate the possibility that an ignored agent causes a collision. The empirical results show this effect directly: safe-scenario rate drops from 100% at N=3 to 86% at N=16 and 85% at N=256. A formal or at least quantitative characterization of this degradation is needed if the paper is to claim scalability with guaranteed safety.","section":"Section II-A and Section IV-D"}],"minor_comments":[{"comment":"The experimental setup says 'all algorithms over 100 distinct initial conditions across 5 seeds', but Section IV-D.2 and Table II use only '20 distinct initial conditions'. Please state the number of seeds for the large-scale experiments as well.","section":"Section IV-C"},{"comment":"The assumptions A1-A4 from [23] are referenced but never listed. Since the viscosity solution result in Eq. (8) depends on them, the paper should at least state them in an appendix and comment on their satisfaction for the considered dynamics.","section":"Section II-B"},{"comment":"The notation in the loss function is inconsistent: H(t, x̂) is defined with x̂ but the second line uses H(ˆok, tk) and H(ˆo, t). Please standardize the argument order.","section":"Eq. (10)"},{"comment":"The phrases 'strict safety guarantees' and 'formal safety guarantees' are used despite the reported safety rates being below 100% in several configurations (e.g., 98.3% safety rate at N=16 in Table I). Consider distinguishing between a certified guarantee and an empirical safety rate.","section":"Abstract and Section I"},{"comment":"The heatmap caption says 'with respect to the ego agent's position coordinates', but the text does not specify the fixed states of the other agents beyond their positions/velocities. Please clarify what is being plotted and how the goal location is encoded.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is an empirical systems paper with a strong engineering contribution, but the formal safety-guarantee framing is not backed by the presented theory. The gaps identified in the major comments are not simply missing citations; they require either new analysis (e.g., a consistency/bounded-mismatch argument, verification of the PINN value function) or a deliberate reframing of the paper as a heuristic method. Given the large experimental component and the reasonable performance, I do not see this as an outright rejection, but the revision should be substantial. I would also flag that the paper leans on self-authored references ([24], [29], [33]) for core elements; the reader should verify that those references indeed contain the stated results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what you should know: this is a solid engineering paper with a real theoretical gap. The central empirical claim — a single PINN value function trained on a 3-agent cluster, deployed per-agent with reachability-based neighbor selection, gives good safety-cost trade-offs at scale — is supported by the simulations. But the formal safety guarantee advertised in the abstract does not hold as stated. The stress-test note is correct: the gap between centralized training and decentralized execution is not bridged. The local safety certificate applies to the joint-optimal trajectory of one cluster, but the actual trajectories are composed of each agent's own local minimizer, which are not jointly optimal for any single cluster. Homogeneity alone doesn't fix that. You need a fixed-point argument or a bounded-mismatch lemma, and none is given. The drop from 100% safe scenarios at 3 agents to 86% at 16 and 85% at 256 is consistent with this gap, though other factors (density, approximation error) could also contribute.\n\nWhat's genuinely new: the specific combination of PINN-based epigraph value learning for multi-agent SC-OCP, the HJ-reachability neighbor selection, and receding-horizon execution. The neighbor-selection ablation is the strongest part — it cleanly shows that safety-value-based selection beats nearest and random. The experiments are run with multiple seeds, report standard errors, and include a scaling study to 256 agents. The paper is honest enough to admit in the conclusion that approximation-error quantification is future work, which somewhat undercuts the earlier formal language.\n\nOther soft spots: the assumptions A1–A4 from Altarovici et al. are referenced but never stated, so the viscosity-solution claim is unverifiable. No code or data is provided, so reproducing the tables is not possible. The paper also overstates the 'globally consistent solution' language in Section II-A without proof.\n\nIs this a serious paper? Yes. The problem is important, the empirical study is careful, and the gap is addressable. A referee should ask for: explicit assumptions, a decentralization consistency argument (or an honest claim that safety is empirical only), an ablation against the single-agent prior framework, and code release. Then this could be a strong contribution. For now, treat the safety guarantees as empirical performance, not verified guarantees.\n\nI would accept it for peer review. The reading group will have a good discussion about the decentralization gap.","headline":"A practically useful decentralized PINN framework for multi-agent SC-OCP, but the safety guarantee is not actually proven — the gap between centralized training and decentralized execution is real and unaddressed.","tokens_in":13083,"tokens_out":2084,"would_cite":true,"duration_ms":17289,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single physics-informed neural network, trained on just three agents, can co-optimize safety and performance for swarms of hundreds.","keywords":["multi-agent safe control","physics-informed neural network","state-constrained optimal control","Hamilton-Jacobi reachability","decentralized policy","epigraph reformulation","collision avoidance","scalable autonomy"],"falsifier":"Run the 256-agent experiment with more random seeds and log every collision; if a nontrivial fraction of collisions involve at least one pair of agents that were not selected as neighbors in the receding-horizon update, the locality assumption is violated. Alternatively, construct a head-on crossing where two non-neighbor agents both converge on the ego agent's future position from outside the observation radius; if the ego agent collides, the claimed globally consistent safety is falsified.","tokens_in":12145,"feed_emoji":"🤖","tokens_out":4047,"duration_ms":31758,"temperature":0.7,"pith_summary":"The paper claims that the hard problem of multi-agent state-constrained optimal control can be split: train one value function on a small homogeneous system, then deploy it independently on each agent using only local observations. The value function is the solution of an epigraph-reformulated Hamilton-Jacobi-Bellman PDE, approximated by a physics-informed neural network with the terminal safety boundary hard-coded. During deployment, agents pick their most safety-critical neighbors using a pairwise reachability value function, then execute the policy in a receding horizon. If correct, this shows that safe, low-cost navigation can scale to hundreds of agents without retraining, because observation size is fixed and each agent runs the same policy.","feed_headline":"One neural net, trained on 3 agents, guides 256 safely","feed_subtitle":"MAD-PINN decentralizes a physics-informed optimal-control value function, giving near-100% collision-free navigation at scale.","key_machinery":"The load-bearing object is the epigraph-reformulated auxiliary value function V̂(t, o, z): a function of state and an auxiliary cost-bound z that is the unique viscosity solution to an HJB PDE (equation 8). It is approximated by a physics-informed neural network whose output is structured as terminal condition + (T−t)·residual, so the safety terminal boundary max(φ(o)−z, g(o)) is satisfied exactly. A second learned object, the pairwise safety value function from HJ reachability, ranks neighbor risk and selects the fixed-size neighbor set. Policy synthesis inverts V̂ via binary search over z to get the minimal admissible cost, then extracts the feedback control from the Hamiltonian minimizer.","core_discovery":"The central claim is that a single distributed value function — trained on a three-agent system — can be deployed in a decentralized manner to any number of homogeneous agents and still co-optimize safety and performance. The paper argues this works because the decentralized SC-OCP has fixed observation size, making the problem dimension independent of total agent count; the epigraph reformulation encodes both cost and safety in one auxiliary value function; and the reachability-based neighbor selection prioritizes safety-critical interactions. Experiments report near-100% safety rates up to 16 agents and above 96% safety rates at 256 agents, with cumulative cost increasing only proportional","pith_inferences":["The paper's locality assumption — that ignoring agents beyond the selected n is safe — is argued by homogeneity and fixed observation size, but no formal error bound is given; a natural next step is a certificate that bounds safety-violation probability as a function of density and n.","If the method transfers to heterogeneous agents or learned dynamics, it would effectively turn multi-agent SC-OCP into a per-agent optimal-control problem with a learned interaction-aware cost, which could apply to warehouse robotics and aerial swarms.","The reported degradation from 100% safe scenarios at 3 agents to 85% at 256 suggests a testable scaling law: one could measure how safe-scenario percentage decays with agent density and use that to pick the required neighbor count for a target safety level."],"forward_implications":["If the value function truly generalizes across agent count, then adding agents to a swarm costs no new training: the same network and the same local observation code handle any N.","Because neighbor selection is based on reachability rather than raw distance, agents anticipate collisions from fast-approaching or velocity-aligned neighbors, not just nearby ones.","The receding-horizon execution makes the policy robust to model mismatch and sensor noise, since the neighbor set and optimization are recomputed online.","Safety and performance are unified in one scalar trade-off z: the agent picks the smallest cost bound that still avoids the unsafe set, so conservatism is controlled not by reward weights but by the constraint structure itself."],"fun_headline_variants":["Train on 3 agents, guide 256 safely","Decentralized physics-informed control scales from 3 to 256 agents","Safety and performance at scale with a neural net trained on 3 agents","Neural net for multi-agent control: train on 3, run on 256 safely"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The framework assumes that interactions with agents outside the chosen neighbor set can be ignored without compromising safety; no theorem or bound is provided to guarantee this in dense crowds.","fun_headline_variants_meta":{"raw":{"variants":["Train on 3 agents, guide 256 safely","Decentralized physics-informed control scales from 3 to 256 agents","Safety and performance at scale with a neural net trained on 3 agents","Neural net for multi-agent control: train on 3, run on 256 safely"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000979,"raw_usage":{"total_tokens":3998,"prompt_tokens":754,"completion_tokens":3244,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":3177}},"tokens_in":498,"tokens_out":3244,"duration_ms":18061,"temperature":1.0,"reasoning_tokens":3177,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T14:38:05.535781+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the 256-agent experiment with more random seeds and log every collision; if a nontrivial fraction of collisions involve at least one pair of agents that were not selected as neighbors in the receding-horizon update, the locality assumption is violated. Alternatively, construct a head-on crossing where two non-neighbor agents both converge on the ego agent's future position from outside the observation radius; if the ego agent collides, the claimed globally consistent safety is falsified.","supporting_citations":[],"review_version":1}