{"id":"ac7595ba-46a3-4641-97b2-57852377c0f2","arxiv_id":"2505.11663","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A multi-robot framework combines clarity-based exploration targets with a robust scheduler that shares one mobile charging station, tested in simulation and on quadrotor hardware.","lead":"This paper presents a system that lets teams of flying robots explore changing environments while sharing one mobile charging station. It combines an information-aware exploration planner with a scheduler that keeps the robots from running out of battery or colliding at the charger.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Energy-feasibility guarantee rests on an unvalidated worst-case battery model; Theorem 1 only holds if f_i^e is a true upper bound, which the constant e_dot = -0.667 hardware model does not establish.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the scheduler's feasibility guarantees are only as strong as the battery discharge model. This is the right focus because Theorem 1 and the RmeSch algorithm reduce to checking conditions computed from f_i^e; if that model is not truly worst-case, the headline guarantee of persistent, energy-feasible operation is unsupported. The paper does present real hardware validation and public code, which count as genuine evidence, but the experiments use a single constant discharge rate with no uncertainty quantification and near-zero final SoC margins. Other concerns, such as the asserted clarity-TISD optimality in Section 4.1 and the Q=0 special case in genTISD, are real but secondary: they affect the quality of exploration and one limiting case, whereas the battery model directly undermines the paper's main formal claim. The appropriate verdict remains CONDITIONAL: the framework is coherent and plausibly useful, but the worst-case energy model, and therefore the persistence guarantee, needs either an explicit uncertainty bound, a robust reserve computation, or experimental validation across discharge uncertainty before the guarantee can be accepted at face value. No change to the reader's conditional verdict is needed.","tokens_in":24114,"tokens_out":8352,"duration_ms":87938,"concrete_test":"Settle this by re-running the Fig. 5 scenario and the 30-robot scalability case with e_dot scaled by 1.1 and 1.2, and with a state/control-dependent discharge term (e.g., an added hover or aggressive-maneuver penalty), while keeping RmeSch unchanged; record the minimum SoC and count violations of (23a). If any run violates the constraint, the claimed guarantee requires an explicit upper bound on |f_true - f_model| or an added reserve margin before it is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central persistence guarantee (23a) is enforced by gware/eware using remaining flight time T_lF, reserve energy e_i,res from (30)-(31), and the Reserve SoC Condition (38). Each quantity is computed by integrating f_i^e from Eq. (1). Theorem 1's proof is conditional on the Gap Flag (35) and Reserve SoC (38) checks passing in every iteration, so if f_i^e underestimates true discharge, a robot can cross e_min before reaching the charger and the 'guarantee' fails in practice. Eq. (1) labels f_i^e as 'worst-case battery discharge dynamics,' but the paper gives no uncertainty set, no conservatism argument, and no validation protocol for that worst-case property. The hardware and simulations use a constant e_dot = -0.667 with no error bars, while the uncertainty-aware reserve (29) covers only position uncertainty of the mobile charger, not discharge-rate uncertainty. The 30-robot scalability run reports returns at (3 +/- 1)% SoC, i.e., almost no margin. Thus the persistent-feasibility claim is established only conditional on a model whose worst-case validity is asserted rather than demonstrated; this is the load-bearing weak point of the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes mEclares, a framework for persistent multi-robot exploration in stochastic spatiotemporal environments, combining a clarity-based target information spatial distribution (genTISD) with ergodic trajectory optimization and an energy-aware scheduler (RmeSch) for a team of rechargeable robots sharing a single mobile charging rover. The authors derive multi-agent clarity dynamics, construct a TISD from the time required to raise each cell's clarity to a target value, and generate candidate trajectories that track a portion of the ergodic plan before returning to a rendezvous point. RmeSch commits or rejects these candidates through gap-aware (gware) and energy-aware (eware) checks, and a fail-safe maneuver is specified for central-node failure. The paper claims feasibility guarantees for the minimum state-of-charge and return-gap constraints, and validates the approach in simulation and on quadrotor hardware with a mobile charging rover; code and an experiment video are released.","tokens_in":24463,"tokens_out":13979,"duration_ms":137758,"significance":"If established, the framework would be a practically useful integration: it supports general nonlinear robot dynamics, a shared mobile charger, online replanning without preplanned schedules, and fail-safe operation under central-node failure, with real hardware experiments and released code. The modular separation between a clarity-driven ergodic planner and an energy scheduler is attractive, and the open-source implementation is a strength. However, the manuscript currently does not establish the claimed link between TISD matching and clarity-deficit minimization, the energy-feasibility guarantee depends on an unvalidated worst-case discharge model, and the reserve-energy formula in Eq. (31a) has a sign error that, as written, makes the energy check vacuous. These are significant but fixable issues, so the appropriate outcome is major revision rather than rejection.","major_comments":[{"comment":"The reserve energy is defined as e_i,res_j = e_i(t_i^f) - e_i(t_i^0), where t_i^0 and t_i^f are the initial and final times of the landing maneuver from the rendezvous point to the worst-case charging position. For any discharging battery, including the constant ˙e = -0.667 used in Sections 8.1 and 8.3, this quantity is negative. The Reserve SoC Condition (38), e_i(t) > e_i,res_j for all t in [t_j, t_i_{j,C}], is then satisfied trivially by every nonnegative trajectory, so eware (Algorithm 4) cannot reject any candidate. The quantity that must appear in (38) is the energy cost of the landing maneuver, e_i(t_i^0) - e_i(t_i^f), which is positive. Because Theorem 1's minimum-energy guarantee relies on (38), and Table 2 distinguishes Baselines 4 and 5 by the action of eware, this sign error is load-bearing. Please correct the sign and re-run or re-verify the affected simulation and hardware claims.","section":"Section 7.2.2, Eq. (31a); Section 7.2.6, Eq. (38)"},{"comment":"The feasibility guarantee is stated as holding 'for all t' and for all robots, but its energy part rests on f_i^e being a worst-case model of battery discharge, and this property is never established. Eq. (1) labels f_i^e 'worst-case battery discharge dynamics,' yet no uncertainty set, no conservatism argument, and no validation protocol are given. The reserve energy in (31), computed via the 95% confidence ellipse in (29), covers only position uncertainty of the mobile charger, not discharge-rate uncertainty. The experiments use the constant ˙e = -0.667 without error bars (Sections 8.1 and 8.3), and the 30-robot scalability result reports returns at (3 ± 1)% SoC, leaving almost no margin. If the true discharge rate is higher than modeled, a robot can cross e_min before reaching the charger, contradicting (23a). Please either provide a calibrated upper bound on discharge with an uncertainty model and feed it into the reserve computation, or explicitly state the guarantee as conditional on the model being an upper bound and add a safety margin in experiments.","section":"Section 7.3, Theorem 1; Eq. (1); Eq. (29)"},{"comment":"The claim that ergodic search with a clarity-based TISD 'naturally minimizes the mean clarity deficit (21)' is not established. Problem (5) minimizes the Sobolev norm between the time-averaged visitation distribution and φ, whereas (24a) minimizes the rectified terminal clarity deficit; these are different objectives. genTISD sets φ_p proportional to ΔT(q_p(t_k), \\bar q_p) and normalizes (Algorithm 1, lines 5-8), which is a heuristic allocation of sensing time, and the nonlinear clarity dynamics (17) mean that proportional time allocation does not generally equalize terminal deficits. The evaluation in Section 8.2 then uses the same mean clarity deficit (21) that defines the TISD, so the comparison partially measures self-consistency rather than an externally grounded information objective. Please provide a formal statement with assumptions under which matching φ minimizes (21), or present genTISD as a heuristic and include a clarity-aware baseline not derived from (21).","section":"Section 4.1; Algorithm 1; Section 8.2"},{"comment":"The induction proof mostly restates the acceptance checks of gware and eware instead of proving the claimed implications. The arrow from the gap-flag inequalities (35) to |t^{i1}_{m1} - t^{i2}_{m2}| > Tδ on [t_j, t_{j,R}) is asserted without an explicit mapping from remaining flight times to actual return times under the candidate and committed trajectories. In the central-node-failure case, the statement that a robot retains enough energy because its previously committed trajectory was feasible does not account for the energy consumed during the idle interval (ret_i - 1)Tδ in Algorithm 5, nor for the time shift in (40); the original trajectory was feasible from t_{j-1}, not from t_j plus an idle delay. Please provide a complete induction with explicit return-time schedules and energy integrals over the full horizon, or state a weaker theorem that matches what is actually proven.","section":"Appendix A.2, proof of Theorem 1"}],"minor_comments":[{"comment":"The 'return RmeSch(...)' branch is a recursive call rather than a blocking wait; when messages are missing this is either an infinite loop or undefined behavior. Rewrite as an explicit wait-and-retry procedure with a timeout.","section":"Algorithm 2, line 3"},{"comment":"The symbol 'li' is not defined, and the fail-safe logic uses ret_i^j inconsistently across Algorithm 2 and Algorithm 5. Please define all return indices and make the notation uniform.","section":"Algorithm 2, line 8; Algorithm 5"},{"comment":"The method name is spelled RmeSch, meSch, and Rmesch at different points (Abstract, Sections 4.2, 7.2, 8.3, Table 2). Standardize the name throughout the manuscript.","section":"Throughout"},{"comment":"R(X) is used as a scalar measurement-noise variance for a single cell, but R(X) is defined in (14) as an N by N matrix. Use a cell-specific scalar, for example R_p(x_i), to avoid an ill-typed expression.","section":"Section 3.2, Eq. (20b)"},{"comment":"The text states that the minimum energy requirement is zero in the hardware experiments, while the 30-robot simulation returns at (3 ± 1)% SoC. Please clarify the relationship between e_min, the reserve energy e_i,res, and the reported margins.","section":"Sections 8.1 and 8.3.1"},{"comment":"The pseudocode writes G_l as (T_l^F - T_R - T_E) > l(Tδ), while Eq. (27) writes G_l = T_l^F > (T_R + T_E + lTδ). They are algebraically equivalent, but presenting the same condition in different forms is confusing; use one form consistently.","section":"Algorithm 3, line 4; Eq. (27)"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Eq. (31a) is serious enough that I would ask the authors to confirm whether the released code implements the positive reserve cost. If the code uses e_i(t_i^0) - e_i(t_i^f), the paper's Eq. (38) and all claims about eware need revision; if the code uses the paper's formula, the energy-violation results in Table 2 need scrutiny. The manuscript is otherwise within scope for a robotics journal, and the incremental novelty over the authors' two prior conference papers is adequately disclosed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know about this paper because it ships code, hardware results, and a coherent framework for persistent multi-robot exploration with a single mobile charger. The new pieces are the multi-agent clarity-based TISD construction (Algorithm 1) and the fail-safe scheduling that handles central-node failure. Those are real contributions. The clarity dynamics derivation (17) is correct, the scheduler logic is sound under its assumptions, and the hardware experiments are genuine. Public code and data help.\n\nThe soft spots are real but not fatal. The persistence guarantee in Theorem 1 leans on a worst-case battery discharge model f_i^e, but the experiments use a constant discharge rate e_dot = -0.667 with no uncertainty bound or validation protocol. The reserve energy calculation covers position uncertainty of the charger but not discharge-rate uncertainty. The 30-robot simulation returns at 3 ± 1% SoC, which is almost no margin. So the 'guarantee' is conditional on a model whose worst-case property is asserted, not established. Theorem 1's proof is mostly a recursive feasibility restatement of the algorithm's own checks; it's not a deep result, though it's not wrong either. And the evaluation metric (mean clarity deficit) is the same quantity used to build the TISD, so the baseline comparison partly measures how well the planner tracks its own target rather than an externally grounded objective.\n\nFor a systems paper, this is solid. The authors are honest about several limitations, and the approach is modular enough that others can use RmeSch as a low-level filter. The main risk is overclaiming the word 'guarantee' without validating the worst-case battery model. A serious referee should ask for that validation, or at least a clearly-stated assumption with conservatism margin, before accepting the persistence claim at face value.\n\nI'd bring this to a reading group to discuss the gap between formal guarantees and real-world validation. I'd cite it if working on energy-aware multi-robot scheduling. It deserves peer review.","headline":"Real hardware and a coherent framework, but the persistence guarantee rests on an unvalidated worst-case battery model and the TISD metric is partly self-referential.","tokens_in":24981,"tokens_out":2916,"would_cite":true,"duration_ms":28690,"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":"mEclares claims that a team of rechargeable quadrotors can keep gathering information indefinitely while sharing one mobile charger, with formal guarantees on battery floors and staggered returns that survive loss of the central…","keywords":["informative path planning","ergodic search","clarity","energy-aware scheduling","multi-robot coordination","persistent missions","mobile charging station","stochastic spatiotemporal environments"],"falsifier":"Run RmeSch with a battery that drains faster than the modeled worst case in any interval (for instance, $\\dot{e} = -0.9$ where the model says $-0.667$) and record state of charge at landing; if the minimum-SoC constraint (23a) is violated before touchdown, or if a scripted central-node dropout ever brings two robots to the charger less than $T_\\delta$ apart, then the feasibility guarantee as stated fails.","tokens_in":23938,"feed_emoji":"🚁","tokens_out":12828,"duration_ms":114193,"temperature":0.7,"pith_summary":"This paper argues that a team of rechargeable robots exploring an environment whose information content decays unpredictably can keep operating forever with a single mobile charging station, and that the guarantees hold in practice, not just in simulation. For exploration, it builds the target density that ergodic search must match out of clarity, a bounded information measure whose decay rate is set by the environment's own process noise, so regions that are losing information fastest attract the most sensing effort. For persistence, it contributes RmeSch, an online scheduler in which each robot proposes a trajectory that follows a slice of its exploration plan and then returns to the charger's predicted position, and only proposals that keep a battery reserve and a minimum gap between charging visits are committed. The paper's Theorem 1 states that, if the fleet is no larger than a bound derived from the least-charged robot's flight time, the battery-floor and exclusive-access constraints hold for all future time, and a fail-safe protocol preserves them even if the coordinating central node fails. If these claims hold, long-term environmental monitoring no longer needs preplanned schedules, static or dedicated charging stations, or a continuously reliable central computer.","feed_headline":"One mobile charger keeps quadrotor teams exploring indefinitely","feed_subtitle":"A clarity-driven planner plus fail-safe scheduling keeps batteries above minimum and charging conflict-free.","key_machinery":"The load-bearing object is the clarity metric $q = 1/(1+P)$ with multi-robot dynamics $\\dot{q} = (1-q)^2 \\sum_{i\\in R} C(x_i)^2/R(x_i) - Qq^2$ (Eq. 17), a bounded $[0,1]$ information measure whose closed-form trajectory (Eq. 18) yields the time-to-target function $\\Delta T(q_0, q_1)$ (Eq. 19). genTISD turns this function into the ergodic target distribution, so the planner's priorities come from the environment model instead of a hand-tuned density. On the scheduling side, the carrying mechanism is the sorted list of remaining flight times feeding gap flags $G_l = T_F^l > T_R + T_E + lT_\\delta$, the reserve energy computed from an EKF prediction of the charger's 95% confidence ellipse, and the gware/eware commit checks; these produce the fleet-size bound in Lemma 1 and the inductive feasibility argument of Theorem 1. The fail-safe machinery is a stored return index per robot, which converts a dead central node into a deterministic staggered return.","core_discovery":"The paper's central claim is that persistent multi-robot information gathering in stochastic spatiotemporal environments reduces to two constructions that plug together. The first sets the ergodic target density at each cell proportional to the time needed to raise that cell's clarity from its current value to the desired target, normalized over the domain (genTISD); since clarity obeys $\\dot{q} = (1-q)^2 \\sum_{i\\in R} C(x_i)^2/R(x_i) - Qq^2$ and admits a closed-form inversion $\\Delta T(q_0, q_1)$, the target density concentrates automatically where information decays fastest or the desired quality is highest. The second lets every robot head for the predicted position of a single mobile charging station, holding in reserve the energy required to reach the worst point inside the 95% confidence ellipse of that prediction, and commits trajectories only when the gap flag $T_F^l > T_R + T_E + lT_\\delta$ and the reserve state-of-charge check both pass. Theorem 1 asserts that with at most $N^* = 1 + \\lfloor (T_{F,0}^{1'} - T_R - T_E)/T_\\delta \\rfloor$ robots and valid initial checks, the minimum-energy and minimum-gap constraints hold for all time, and that if the central node dies, rank-based idle times followed by replay of the last committed trajectory preserve both constraints. The paper presents multi-quadrotor hardware experiments with a ground-based charging rover as evidence that the pipeline runs in real time.","pith_inferences":["The claim that ergodic search with the clarity-derived TISD 'naturally minimizes' the mean clarity deficit is argued by motivation (Section 4.1) rather than proven; if a formal equivalence were established, the same $\\varphi_p = \\Delta T(q_p, q_p^*)$ construction would transfer directly to any scalar field with known stochasticity, such as ocean or atmospheric variables.","The constant worst-case discharge rate used in experiments ($\\dot{e} = -0.667$ with no uncertainty margin) is the weakest seal in the guarantee chain; validating the scheduler with an online, conservative discharge bound under wind or payload variation would turn the energy-feasibility theorem from conditional into demonstrated.","Because the fail-safe protocol only requires a rank and a stored trajectory, the same scheduling logic ports to fully decentralized or intermittently connected networks with little change, although the paper lists that direction as future work.","The charging gap $T_\\delta$ acts as a scheduling quantum: Lemma 1 sets the maximum team size roughly as $1/T_\\delta$, so a mission needing more robots must shorten charging-plus-buffer time or accept a smaller fleet; the paper does not discuss this trade as a design parameter."],"forward_implications":["Persistence without preplanning: under Theorem 1, any committed trajectory is energy-feasible and charger-conflict-free for all time, as long as the fleet respects the Lemma 1 size bound.","Ergodic search gains a model-derived target density: with process noise $Q_p$ known per cell, clarity-based genTISD replaces uniform or hand-tuned densities and, per the paper's Section 4.1 argument, drives down the mean clarity deficit as a direct effect of ergodicity.","No static charging infrastructure: the mobile charger follows the geometric center of the team's nominal trajectories, and RmeSch runs as a planner-agnostic module that the paper demonstrates with both ergodic and non-ergodic (Lissajous) coverage planners.","Central node failure is survivable: each robot needs only its latest committed trajectory and its return index, so exclusive charging access and the battery floor survive communication loss beyond the decision interval.","Fleet size can change online: new robots can be added whenever the least-charged robot's flight time satisfies the Remark 2 condition, and failed robots can be dropped from gap-flag checks without breaking the guarantees for the rest."],"supporting_citations":[{"why":"Defines the clarity metric and its dynamics that the target-density construction and the closed-form $\\Delta T$ rest on.","marker":"D. R. Agrawal and Panagou (2023)"},{"why":"Supplies the ergodicity metric $\\|\\cdot\\|_{H^{-(s+1)/2}}$ that the clarity-driven target distribution is matched against.","marker":"G. Mathew and Mezić (2011)"},{"why":"Provides the Fourier-based ergodic trajectory generation used as the nominal planner.","marker":"Dressel and Kochenderfer (2019)"},{"why":"Prior clarity-driven single-robot planner that serves as the baseline whose simultaneous charging visits RmeSch fixes.","marker":"Naveed, Agrawal, Vermillion, and Panagou (2024b)"},{"why":"Prior multi-agent scheduler (meSch) that this work extends with the fail-safe planner and the fuller feasibility analysis.","marker":"Naveed, Dang, Kumar, and Panagou (2024)"},{"why":"Staggered-deployment baseline that achieves charging exclusivity by delaying launches; RmeSch improves on it by launching all robots at once.","marker":"Bentz et al. (2018)"},{"why":"Control-barrier-function scheduling baseline with simplified dynamics; the comparison that motivates support for general nonlinear models.","marker":"Fouad and Beltrame (2022)"},{"why":"Supplies the nonlinear quadrotor dynamics used for candidate trajectory generation and the hardware experiments.","marker":"Jackson et al. (2021)"},{"why":"Provides the EKF prediction equations used to estimate the charger's position and covariance, from which the reserve energy is computed.","marker":"Gelb et al. (1974)"},{"why":"The recursive-feasibility proof of Theorem 1 is explicitly modeled on this work's proof structure.","marker":"D. Agrawal et al. (2023)"}],"fun_headline_variants":["Lone mobile charger keeps multi-robot exploration going indefinitely","Fail-safe scheduling with one mobile charger enables nonstop robot missions","Clarity-driven adaptive search plus one charger sustains persistent exploration","Single mobile charger powers long-term multi-robot information gathering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The energy-feasibility guarantee rests entirely on the worst-case battery discharge model $f_i^e(e_i)$ in Eq. (1) really being an upper bound on how fast each battery can drain, and the experiments use a constant rate $\\dot{e} = -0.667$ with no uncertainty margin, so a real battery that drains faster than modeled could hit the minimum state of charge before it reaches the charger.","fun_headline_variants_meta":{"raw":{"variants":["Lone mobile charger keeps multi-robot exploration going indefinitely","Fail-safe scheduling with one mobile charger enables nonstop robot missions","Clarity-driven adaptive search plus one charger sustains persistent exploration","Single mobile charger powers long-term multi-robot information gathering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000418,"raw_usage":{"total_tokens":2219,"prompt_tokens":1076,"completion_tokens":1143,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":1074}},"tokens_in":692,"tokens_out":1143,"duration_ms":11834,"temperature":1.0,"reasoning_tokens":1074,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:51:13.931563+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run RmeSch with a battery that drains faster than the modeled worst case in any interval (for instance, $\\dot{e} = -0.9$ where the model says $-0.667$) and record state of charge at landing; if the minimum-SoC constraint (23a) is violated before touchdown, or if a scripted central-node dropout ever brings two robots to the charger less than $T_\\delta$ apart, then the feasibility guarantee as stated fails.","supporting_citations":[{"cited_title":"\\ Kochenderfer, M J","cited_arxiv_id":null,"evidence_quote":"Provides the Fourier-based ergodic trajectory generation used as the nominal planner."},{"cited_title":", Hoang, T","cited_arxiv_id":null,"evidence_quote":"Staggered-deployment baseline that achieves charging exclusivity by delaying launches; RmeSch improves on it by launching all robots at once."},{"cited_title":", Tracy, K","cited_arxiv_id":null,"evidence_quote":"Supplies the nonlinear quadrotor dynamics used for candidate trajectory generation and the hardware experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the EKF prediction equations used to estimate the charger's position and covariance, from which the reserve energy is computed."}],"review_version":1}