{"id":"8c94c2dc-d365-47b7-a614-b2088251871b","arxiv_id":"2507.21338","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A bimodal (drive-and-fly) robot explores unknown environments by using a Monte Carlo Tree Search planner that selects both the next viewpoint and the locomotion mode under energy and time limits.","lead":"This paper presents a planning framework that lets a hybrid ground-and-air robot automatically choose when to roll or fly while exploring unknown buildings under a fixed energy and time budget. It could make search-and-rescue robots last longer by saving battery power for the parts of a mission that need flight.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The constant-power energy model of Eq. (4) is load-bearing: simulation metrics are computed from the same model that BM-MCTS optimizes, and the lone real-world run lacks baselines and independent battery measurements.","rationale":"The paper's central claim is conditional on the energy/time cost model of Eqs. (3)-(4) resembling the physical platform. All simulated evidence for adaptivity (Fig. 7e-g, Fig. 8a-b, and Tabs. 1-2 in the supplement) is computed using the same constant-power model that BM-MCTS optimizes, so it cannot distinguish good planning from self-fulfilling prophecy. The single real-world run (Sec. VII-B) is a valuable existence proof that the system can fly over stairs, roll, and return, but it has no baseline and does not report predicted vs. measured energy per segment, so it does not resolve the model question. I checked for internal inconsistencies: the M=(T+A)/2 guidance-path approximation in Eq. (8) is acknowledged, and the paper's own limitation section (Supp. Sec. 11) admits 'simplified terrain analysis', which is directly relevant. These admissions support, not weaken, the conditional verdict. A concrete computational experiment, perturbing the energy model toward state-dependent power and re-running the core comparison, would determine whether the reported modality trade-offs and coverage gains are robust. If they vanish under perturbation, the model is load-bearing; if they persist, the central claim is strengthened. This is the single most decisive check, short of instrumented hardware trials with baselines. Overall the reader's CONDITIONAL verdict is appropriate; my analysis does not change it.","tokens_in":19466,"tokens_out":5445,"duration_ms":66588,"concrete_test":"Re-run the supplementary Sec. 8.2 Case 1 (E300, T400) comparison with a modified energy model that multiplies P_M by a terrain factor (e.g., 2.5x for segments crossing stairs or modality transitions) and optionally adds a fixed takeoff/landing energy penalty; keep all other components fixed. If the BM-MCTS modality ratio (ground vs. air time) or final coverage changes by more than 10% relative to the values reported in Tabs. 1-2, the constant-power assumption of Eq. (4) is load-bearing and the central claim requires hardware validation with a battery logger and an aerial-only baseline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (4) models energy as E = P_M * T(Pi, Pj, M) with P_A/P_T = 7, so every simulated remaining-energy and remaining-time reading in Sec. VII-A and in the supplementary comparisons (Tabs. 1-2) is generated by the very cost model the planner optimizes. The paper's central claim that BM-MCTS adapts modality to energy/time budgets and outperforms an aerial-only baseline under energy shortage therefore rests on the realism of that model. The model ignores state-dependent power: climbing stairs, terrain resistance, and takeoff/landing transitions are not reflected, even though the bimodal motion planner (Sec. VI) explicitly uses aerial primitives for stair climbing and edge crossings. Eq. (7) assigns the modality of the arrival viewpoint to the entire segment, so a terrestrial-to-terrestrial segment that actually requires a brief flight is charged at terrestrial power. The real-world experiment (Sec. VII-B) reports one run, no baseline, and plots energy/modality curves without comparing predicted vs. measured segment consumption, so it does not independently validate the model. If real power draw is state-dependent, the planner's modality choices could be systematically wrong on hardware, and the reported coverage advantage would not transfer.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hierarchical autonomous exploration framework for terrestrial-aerial bimodal vehicles (TABVs). The system generates bimodal viewpoints for frontier clusters, uses an extended Monte Carlo Tree Search (BM-MCTS) to select viewpoint sequences and locomotion modalities under energy and time budgets, and integrates an improved bimodal motion planner with terrain perception. The claimed contribution is that the vehicle adapts its modality to satisfy E_all and T_all constraints, achieving higher coverage than aerial-only baselines under energy shortage and comparable performance under time shortage. The paper reports simulations in multi-story buildings, budget-adaptability experiments, and a single real-world deployment in an underground parking garage.","tokens_in":19742,"tokens_out":3538,"duration_ms":41205,"significance":"If the claims hold, the work is a valuable step toward practical deployment of TABVs for search-and-rescue-style exploration, where energy and time budgets matter. The system is complete: perception, viewpoint generation, decision making, motion planning, and control are integrated, with a real-world demonstration. The supplementary material provides useful ablations: comparisons against FUEL, UAV-only, UGV-only, and UGV-cross baselines, and an analysis of the topo-graph's computational advantage. The problem formulation in Eq. (1)-(2) is clear, and the BM-MCTS design with discounted information gain and exponential penalties is well motivated. However, the evaluation has important weaknesses: the simulation energy/time metrics are generated by the same constant-power model the planner optimizes, the real-world test has no baseline, and several key comparisons appear only in the supplementary rather than the main text.","major_comments":[{"comment":"The constant-power energy model E(P_i, P_j, M) = P_M * T(P_i, P_j, M) is load-bearing: the remaining-energy and remaining-time values reported in Sec. VII-A and in the supplementary comparisons (Tabs. 1-2) are computed from the same model that BM-MCTS optimizes. The model ignores state-dependent power draw, such as stair climbing, edge transitions, and terrain resistance, even though the motion planner explicitly uses aerial primitives for such cases (supplementary Sec. 2, Fig. 4). The real-world experiment in Sec. VII-B reports a single run without a baseline, without independent battery measurements, and without comparing predicted versus measured segment consumption. As a result, the simulated modality-balancing behavior may be an artifact of the cost model rather than physical reality. I request either (a) independent validation of Eq. (4) against logged power data from the platform, or (b) repeated real-world runs comparing the full system against UAV-only and UGV-only baselines, with measured battery consumption.","section":"Sec. V-A, Eq. (4); Sec. VII-B"},{"comment":"Equation (7) charges the entire segment from P_i to P_{i+1} at the power of the modality assigned to the arrival viewpoint, M(P_{i+1}). This means a terrestrial-to-terrestrial segment that actually requires a brief flight (e.g., to climb stairs, as illustrated in Fig. 4 of the supplementary) is charged at terrestrial power, underestimating energy consumption. This bias could systematically favor terrestrial choices in simulation, affecting the reported coverage and remaining-energy values. Please either modify the cost model to account for intermediate modality switches, or provide evidence that such mixed-modality segments are rare enough in the tested scenes to be negligible.","section":"Sec. V-B, Eq. (7)"},{"comment":"The paper reports five or ten trials per condition but does not provide error bars, standard deviations, or statistical significance for the coverage percentages and remaining-energy/time values. For example, in supplementary Tab. 2, Case 1, the coverage difference between UGV-cross (92.1%) and Ours (93.2%) is small, and without variance information it is unclear whether this difference is meaningful. Please report the variance across trials and, where appropriate, statistical tests for the key comparative claims.","section":"Sec. VII-A, Fig. 8; supplementary Tabs. 1-2"},{"comment":"The guidance path in the simulation phase uses the average modality M = (T+A)/2 (Sec. V-C and Eq. (8)), whereas the tree expansion uses actual per-segment modalities (Eq. (7)). The supplementary (Sec. 5) acknowledges this inaccuracy and argues that the UCB selection policy progressively corrects it, but no empirical evidence is given that the correction converges within the iteration counts used in Sec. VII-A (e.g., ten iterations in supplementary Sec. 8.1). Please quantify the estimation error introduced by the average-modality assumption, for example by comparing the guidance-path estimates against the fully expanded tree costs, or by analyzing sensitivity to the choice of P_{(T+A)/2}.","section":"Sec. V-C, Eq. (8); supplementary Sec. 5"}],"minor_comments":[{"comment":"The main text does not include the baseline comparisons that support the abstract's claim of 'better coverage than an aerial-only baseline under energy shortage and comparable performance under time shortage.' These results appear only in the supplementary (Tabs. 1-2). I recommend citing these tables explicitly in Sec. VII-A or moving a summary of the framework-level comparison into the main text.","section":"Sec. VII-A"},{"comment":"The real-world experiment reports only one run. Please state the number of runs explicitly and, if multiple runs were performed, report the spread of energy/modality trajectories. Also, the energy and modality curves in Fig. 10(b) would be more informative if annotated with the times of takeoff/landing events and with the measured versus predicted remaining energy.","section":"Sec. VII-B"},{"comment":"The time cost model uses max(length/v_M,max, dyaw/omega_M,max). This neglects acceleration limits and turning time during path traversal. This is a simplification, but since the same model is used in simulation and planning, the resulting bias is only relevant to real-world transfer; please mention this limitation explicitly.","section":"Eq. (3)"},{"comment":"The hyperparameter formulas a1 = b1 - log(0.3), b1 = log(10) should be checked for typographical consistency: the main text (Sec. V-A) defines kappa_Er(x) = exp(-a1*x/E_all + b1) and kappa_Tr(x) = exp(-a2*x/T_all + b2). Please clarify the units of b1 and b2 (the offset in the exponent must be dimensionless).","section":"Supplementary Sec. 9.1"},{"comment":"Reference [6] is cited as the basis for the bimodal motion planner, but the extension over [6] is described only qualitatively. Please highlight the specific new capabilities (e.g., online ground segmentation, dual grid maps, modality-aware primitives, ESDF-based edge safety) in a way that lets readers identify the novel components without reading the supplementary.","section":"Sec. II"}],"recommendation":"major_revision","confidential_remarks":"The paper's main-text evaluation is thin; the most convincing comparisons (FUEL, UAV-only, UGV-only, UGV-cross) are in the supplementary. The energy-model validation issue is the principal technical concern. I would like the editor to ensure that the revised version addresses the model-validation request with real data or a clearly stated and justified limitation. The paper is within the journal's scope and the system is impressive, but the current evidence does not yet support the strong claims made in the abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"BM-MCTS is the real new piece: it folds locomotion modality into the MCTS decision under energy and time budgets, and the paper builds a complete exploration system around it, with a hardware demo. That is worth taking seriously. The main soft spot is that the simulation's resource metrics come from the same constant-power model the planner optimizes, so the balancing results are partly self-confirming.\n\nThe paper does several things well. The two coverage strategies (aerial-only vs hybrid) for viewpoint generation are a sensible way to give the planner modality options. The supplementary comparisons are stronger than the main text suggests: they include FUEL as a framework baseline, UAV-only/UGV-only/UGV-cross ablations, and NBV/TSP algorithm comparisons. The UGV-cross ablation is the right control and shows the MCTS's proactive switching buys something over a rule-based 'fly only when necessary' policy. The real-world run, though single, is a genuine demonstration with a 7.2x power ratio and a hard energy budget.\n\nSoft spots, in order. First, Eq. (4) is a constant-power model, and Eq. (7) charges a whole segment at the arrival viewpoint's modality, so a terrestrial segment that actually needs a brief flight for a stair or edge is billed at terrestrial power. Second, the simulation's remaining energy and time are computed with the same model, so Fig. 8's budget-adaptation trends are partially circular. Third, the hardware experiment is one run, no baseline, and no predicted-vs-measured segment comparison, so it does not independently validate the model. These are real limitations, but they don't sink the contribution; the system works end-to-end at least once, and the ablations are consistent with the model. The authors also acknowledge simplified terrain analysis and local resource estimation in the supplementary limitations.\n\nWho this is for: anyone working on energy-aware exploration or multi-modal robots will get something from the MCTS formulation and the ablation design. It deserves a serious referee. I'd send it out, and in review I'd push for independent battery logging, repeated runs, error bars, and a clear separation between model-generated metrics and measured ones. With that added, the conditional becomes a solid accept.","headline":"BM-MCTS is a genuine new module for modality-aware exploration, but the simulation's resource metrics are generated by the same constant-power model the planner optimizes, so independent validation is needed.","tokens_in":20261,"tokens_out":2198,"would_cite":true,"duration_ms":25779,"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":"The paper claims a fly-and-roll robot can choose its mode leg by leg so as to maximize explored information while still returning home within energy and time budgets, and supports the claim with simulations and a real deployment.","keywords":["autonomous exploration","terrestrial-aerial bimodal vehicle","Monte Carlo tree search","energy-aware planning","time-aware planning","viewpoint planning","bimodal motion planning","frontier-based exploration"],"falsifier":"Measure the physical platform's battery draw across repeated stair ascents, edge crossings, and flat rolling in both modalities and compare integrated energy use with $E(P_i,P_j,M)=P_M T(P_i,P_j,M)$. If the actual energy per meter is not approximately constant per modality, or the aerial-to-terrestrial ratio departs materially from 7 on these maneuvers, then the planner's simulated remaining-energy estimates, and the modality switches they cause, would not carry over to hardware.","tokens_in":19267,"feed_emoji":"🤖","tokens_out":8822,"duration_ms":95928,"temperature":0.7,"pith_summary":"The paper claims that a single robot that can both roll and fly can explore an unknown 3-D space under explicit energy and time budgets, if the choice of locomotion mode is made part of the planning problem rather than fixed by hardware. The authors build a hierarchical planner that generates candidate aerial and hybrid viewpoints, uses an extended Monte Carlo tree search to choose both the order of viewpoints and the modality used to reach each one, and sends the chosen path to a bimodal motion planner. They formulate the task as maximizing information gain subject to nonnegative remaining energy and time, with a harder penalty on running out of energy than on running late. In simulation the bimodal planner outperforms an aerial-only baseline under energy shortage, reaching 93.2 percent coverage versus 58.4 percent in the benchmark case, and stays comparable under time shortage. A real platform completes an underground-garage exploration on a 20 Wh budget, so the payoff, if the claim holds, is that practical search-and-rescue exploration can be done by one vehicle instead of a coordinated aerial-ground team.","feed_headline":"Fly-roll robot adapts travel mode to energy and time budgets","feed_subtitle":"A tree-search planner lets a two-mode vehicle cover more ground than flying alone when battery is tight.","key_machinery":"The central object is BM-MCTS, a Monte Carlo tree search extended from the classic selection-expansion-simulation-backpropagation loop. Each tree node is a viewpoint; each root-to-leaf branch is a candidate traverse sequence; the modality used to travel into a node is encoded by that node, so the search decides modality and sequencing at once. The reward per node splits into a process gain, a discounted average of visible frontier-voxel information, and a terminal cost built from exponential penalties on remaining energy and time, and the UCB rule balances exploration against exploitation when choosing children. A grouped TSP over frontier-cluster viewpoints, evaluated on a topological graph with an average-modality cost model, supplies the long-horizon estimate of what it takes to finish and return. The energy model behind the estimates is constant power per modality, $E(P_i,P_j,M)=P_M T(P_i,P_j,M)$, with flying power roughly seven times rolling power.","core_discovery":"On the paper's own terms, the discovery is that resource-constrained exploration of a bimodal vehicle reduces to a search over viewpoint sequences in which every edge of the sequence carries a modality label, and that this joint search is tractable enough to run online. The BM-MCTS tree stores discounted information gain and exponential penalties for low remaining energy and time; a simulation phase estimates the cost of finishing all frontier clusters and returning home by solving a grouped traveling-salesman path on a topological graph, and a pruning rule stops expanding any branch that cannot return with nonnegative energy. The same framework is then coupled to a motion planner that perceives terrain online, builds separate maps for rolling and flying, and applies modality-specific dynamics. The reported consequence is that the robot's mode choices follow the budget: it flies while energy is plentiful and information is high, rolls when energy is scarce, and switches early enough to leave a safety margin for coming back.","pith_inferences":["The same BM-MCTS scheduling could be transferred to other hybrid mobility platforms, for example legged or hopping vehicles with two very different cost regimes, as long as the per-modality cost model is recalibrated; the paper does not test this transfer.","The steep energy penalty versus mild time penalty encodes a particular mission priority: returning alive matters more than finishing on time. Swapping those priorities would change the branches the tree prunes, so the reported behaviors are partly a consequence of that design choice rather than of the search alone.","A natural follow-up is to replace the constant-power energy model with terrain-aware consumption learned from onboard current sensing; the paper's own limitation section notes that resource estimates are locally myopic, and a terrain-sensitive model would address both that and the constant-power assumption."],"forward_implications":["A single terrestrial-aerial vehicle can replace a multi-robot air-ground team for budgeted exploration, avoiding multi-robot SLAM, communication, and coordination overhead.","Energy-limited missions can be planned to prioritize returning with data: the planner stops expanding branches whose remaining energy cannot cover the trip home.","The modality ratio responds to the budget: tighter energy budgets push the vehicle toward rolling, tighter time budgets push it toward flying, and the planned margin leaves room for uncertainty.","Computational cost stays manageable as the number of frontier clusters grows, because the topo-graph makes the simulation-phase path estimates cheap.","The framework works with limited field-of-view sensors by passing the sensor FoV into viewpoint generation, so the same planner can be fitted to different perception hardware."],"supporting_citations":[{"why":"Supplies the previous bimodal navigation and control pipeline that this work extends with terrain perception and modality-aware planning.","marker":"[6]"},{"why":"Defines the frontier-based exploration paradigm used to locate informative regions.","marker":"[7]"},{"why":"One of the bimodal vehicle designs cited to justify the assumed 5–8 times higher aerial power consumption.","marker":"[17]"},{"why":"Provides the submodularity guarantee behind the greedy bimodal viewpoint selection.","marker":"[20]"},{"why":"Another bimodal vehicle design cited for the aerial-to-ground power ratio underlying the energy model.","marker":"[22]"},{"why":"Introduces the Monte Carlo tree search structure that BM-MCTS extends.","marker":"[24]"},{"why":"Supplies the UCB-style selection policy and multi-objective informative planning that BM-MCTS adapts.","marker":"[25]"},{"why":"Serves as the state-of-the-art aerial-only baseline in the framework-level comparison.","marker":"[5]"},{"why":"Provides the lidar-inertial odometry used for onboard localization in the real-world experiment.","marker":"[27]"}],"fun_headline_variants":["Fly-roll robot uses tree search to ration energy and time","Bimodal explorer switches modes to squeeze more coverage","Smart mode choices let aerial-ground robot last longer","Tree search plans flights and drives for efficient mapping","Energy-aware Monte Carlo search guides fly-roll robot"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The plan's value rests on the assumption that a flying robot uses roughly seven times the power of the same robot rolling, at a constant per-second rate, for every maneuver including stairs and terrain edges; if real power draw varies with terrain or speed, the modality choices are optimized for a model rather than for the vehicle.","fun_headline_variants_meta":{"raw":{"variants":["Fly-roll robot uses tree search to ration energy and time","Bimodal explorer switches modes to squeeze more coverage","Smart mode choices let aerial-ground robot last longer","Tree search plans flights and drives for efficient mapping","Energy-aware Monte Carlo search guides fly-roll robot"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1512,"prompt_tokens":849,"completion_tokens":663,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":465,"completion_tokens_details":{"reasoning_tokens":587}},"tokens_in":465,"tokens_out":663,"duration_ms":8153,"temperature":1.0,"reasoning_tokens":587,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:52:24.829204+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the physical platform's battery draw across repeated stair ascents, edge crossings, and flat rolling in both modalities and compare integrated energy use with $E(P_i,P_j,M)=P_M T(P_i,P_j,M)$. If the actual energy per meter is not approximately constant per modality, or the aerial-to-terrestrial ratio departs materially from 7 on these maneuvers, then the planner's simulated remaining-energy estimates, and the modality switches they cause, would not carry over to hardware.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the previous bimodal navigation and control pipeline that this work extends with terrain perception and modality-aware planning."},{"cited_title":"Williams, S","cited_arxiv_id":null,"evidence_quote":"Defines the frontier-based exploration paradigm used to locate informative regions."},{"cited_title":"Zheng, Q","cited_arxiv_id":null,"evidence_quote":"One of the bimodal vehicle designs cited to justify the assumed 5–8 times higher aerial power consumption."},{"cited_title":"Duawlfin: A Drone with Unified Actuation for Wheeled Locomotion and Flight Operation","cited_arxiv_id":"2505.13836","evidence_quote":"Provides the submodularity guarantee behind the greedy bimodal viewpoint selection."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Another bimodal vehicle design cited for the aerial-to-ground power ratio underlying the energy model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the Monte Carlo tree search structure that BM-MCTS extends."},{"cited_title":"Kalantari, T","cited_arxiv_id":null,"evidence_quote":"Supplies the UCB-style selection policy and multi-objective informative planning that BM-MCTS adapts."},{"cited_title":"Ropero, P","cited_arxiv_id":null,"evidence_quote":"Serves as the state-of-the-art aerial-only baseline in the framework-level comparison."},{"cited_title":"Chen and L","cited_arxiv_id":null,"evidence_quote":"Provides the lidar-inertial odometry used for onboard localization in the real-world experiment."}],"review_version":1}