{"id":"589e387f-84c9-4c54-b71d-ad6efd647393","arxiv_id":"2412.19744","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A simulator integrates particle-based water, aerial flight, manipulation, and animal models to train aerial-aquatic robots across air and water.","lead":"Researchers built AAM-SEALS, a photorealistic simulator on NVIDIA Isaac Sim where aerial-aquatic drones with robotic arms can fly, dive, interact with underwater animals, and train reinforcement learning policies. It aims to give robotics labs a safe testbed for cross-medium search, rescue, and aquaculture tasks before building physical robots.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Hydrodynamic fidelity rests on a single unquantified drop-test plot; the abstract's 'quantitative position-tracking' validation is not what Sec. V-A reports.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: particle-based PBD fidelity is asserted but not quantitatively established. My stress-test adds two supporting observations: the abstract's description of the validation ('position-tracking errors') does not match the actual experiment (acceleration curves), and the absence of the rigid-body baseline (which the authors say they integrated) makes the qualitative match non-diagnostic. The paper's other components — photorealistic rendering, RL demonstrations, and the integrated control stack — are plausible contributions, but they do not independently establish hydrodynamic fidelity. Since the missing evidence is in principle obtainable by releasing code/data and running a baseline comparison, the conditional verdict is appropriate. If the concrete test shows PBD does not beat the rigid-body baseline, the verdict should move toward rejection of the high-fidelity claim; if it does, the claim stands. Hence no change to the reader's verdict is needed now.","tokens_in":16432,"tokens_out":2367,"duration_ms":26289,"concrete_test":"Reproduce the Sec. V-A drop in SEALS using both PBD and the rigid-body hydrodynamics baseline, with the simulated AAM's mass, inertia, center-of-gravity, and drop height matched to the real 3D-printed model. Run at least 10 trials per condition, record z-acceleration and position-vs-time, and compute RMSE and time-to-peak deceleration relative to the real IMU data. If PBD does not significantly outperform the rigid-body baseline, or if no quantitative comparison is reported, the high-fidelity claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that particle-based hydrodynamics in SEALS is 'high-fidelity' rests entirely on Sec. V-A, which compares one simulated free-fall-into-water acceleration curve (Fig. 10) with one real-world IMU curve. No error metric, repeat count, error bars, or statistical test is reported. Appendix A states that a rigid-body hydrodynamics baseline was integrated alongside PBD, but no comparison to that baseline is shown, so the qualitative similarity cannot be attributed to PBD rather than to a tuned drag or buoyancy model. The abstract claims validation 'by comparing position-tracking errors across real-world and simulated systems,' but Sec. V-A measures z-axis acceleration over time, not position tracking, and no position-tracking error is reported for the hydrodynamics experiment. In addition, the simulated AAM's mass, inertia, center-of-gravity, and drop height are never matched to the 3D-printed model, leaving the comparison underconstrained. The paper's headline novelty therefore depends on an unverified modeling assumption, and the evidence as presented cannot distinguish PBD from a simpler, parameter-fitted hydrodynamics model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces AAM-SEALS, a simulator built on NVIDIA Isaac Sim for Aerial-Aquatic Manipulators (AAMs) that aims to integrate flight, swimming, and manipulation across sea, air, and land. The system uses position-based dynamics (PBD) via PhysX for hydrodynamics, simplified linear drag for aerodynamics, PID/PD controllers for the vehicle and manipulator, and offers cameras, contact sensors, and RL interfaces. The evaluation claims quantitative validation of hydrodynamic fidelity by comparing simulated and real-world drop tests, demonstrates position-tracking for hovering and cross-medium trajectory following, and reports RL and RLfD training results. The paper also describes a new robot class, a photorealistic environment with aquatic animal models, and an open-source commitment.","tokens_in":16617,"tokens_out":3344,"duration_ms":33017,"significance":"If the validation were made rigorous, AAM-SEALS would be a valuable and timely contribution: it is apparently the first integrated simulator for aerial-aquatic manipulation that spans sea, air, and land, with photorealistic rendering, particle-based hydrodynamics, and learning interfaces. The authors provide a real-world drop-test data point, detailed appendices on PBD and RL hyperparameters, and a stated intention to open-source code and data; these are concrete assets. However, the central fidelity claim currently rests on a single qualitative acceleration-curve comparison, and the paper's own appendix admits that a rigid-body hydrodynamics baseline was integrated but never reported. The contribution is original and the direction is sound, but the evidence presented does not yet support the advertised quantitative validation.","major_comments":[{"comment":"The abstract's central validation claim—'quantitatively validate the fidelity of particle-based hydrodynamics by comparing position-tracking errors across real-world and simulated systems'—does not match the experiment reported in Sec. V-A, which compares one simulated z-axis acceleration-over-time curve with one real-world IMU curve (Fig. 10). No position-tracking error, numerical error metric, error bars, or repeat trials are reported, so the claimed quantitative validation is absent; the abstract and experiment need to be reconciled.","section":"Abstract and Sec. V-A"},{"comment":"Appendix A states that a rigid-body hydrodynamics baseline was integrated into AAM-SEALS alongside position-based hydrodynamics 'to compare the two hydrodynamics models,' yet no baseline comparison appears in the evaluation. Without this baseline, the qualitative similarity in Fig. 10 cannot be attributed to PBD rather than to the linear drag model, buoyancy, or other tuned simulation parameters; the authors should report the baseline comparison or remove the claim that it provides insight.","section":"Sec. V-A and Appendix A"},{"comment":"The real-world and simulated drop tests are underconstrained: the mass, inertia, center of gravity, drop height, and water tank properties of the 3D-printed AAM are not specified as matched simulation parameters. The reported qualitative agreement therefore does not establish predictive fidelity; at minimum, the authors should report these parameters and ideally include a sensitivity analysis to show which parameters drive the observed agreement.","section":"Sec. V-A"},{"comment":"The control evaluation reports position tracking within ±0.015 m, ±0.003 m, and ±0.2 m for X, Y, and Z during hovering, but the PID gains in Eq. (6) and PD gains in Eq. (7) are not reported, and no repeat trials or disturbance conditions are described. Since the adaptive-allocation mechanism in Eq. (4) is central to the claimed robustness to a changing center of gravity, the tracking results need at least gain values and a statement of how many runs they summarize.","section":"Sec. V-B and Eqs. (6)-(7)"}],"minor_comments":[{"comment":"The caption misspells 'Proportional' as 'Propotional' twice; please correct both instances.","section":"Fig. 3 caption"},{"comment":"The caption ends with an incomplete fragment, 'steps'; it should be completed or removed.","section":"Fig. 15 caption"},{"comment":"The sentence 'This system gives SEAL a strong and cutting-edge balance' should read 'SEALS' rather than 'SEAL'.","section":"Sec. IV-B"},{"comment":"The sentence 'All of the objects are equipped with an IMU sensor' is not supported by the reported experiments, which describe only the 3D-printed AAM; please clarify which objects were actually tested.","section":"Sec. V-A"},{"comment":"The reward regions are described as 'outer (distance greater than 1 meter), inner (distance between 1 meter and dt), and success (distance less than dt),' where dt is 10^{-2} m; this leaves the interval (dt, 1 m) ambiguously assigned, so the boundaries should be restated more precisely.","section":"Appendix C"},{"comment":"The Limitations paragraph already acknowledges that Sim2Real transfer is not fully verified; this is appropriate, but it should be cross-referenced with the fidelity claims made in the abstract and Sec. V-A so that readers are not misled.","section":"Sec. VI"}],"recommendation":"major_revision","confidential_remarks":"The paper's abstract significantly overstates what Sec. V-A actually reports. The authors have the ingredients for a sound systems paper—a real drop-test experiment, a PBD-based simulator, and a stated rigid-body baseline—but they need to either add the missing quantitative comparison and baseline evaluation or substantially soften the fidelity claim. I believe this is fixable within the scope of a revision and do not see a case for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Yang and colleagues have built something useful: a single Isaac Sim platform that ties together a quadrotor-with-manipulator model, particle-based hydrodynamics, controllable crab and sea spider models, and RL training pipelines. That combination is new, as far as I can tell from the related work, and the paper is honest about what it does not do — the limitations section explicitly says Sim2Real transfer is not fully verified. The real-world drop test is a genuine attempt at external validation, and the plan to open-source code and data is the right move.\n\nThe soft spots are real but fixable. The abstract says the fidelity of particle-based hydrodynamics is 'quantitatively validate[d] by comparing position-tracking errors across real-world and simulated systems.' That is not what Section V-A reports. The experiment compares one simulated acceleration curve with one real IMU curve from a 3D-printed model. There are no error metrics, no repeat trials, no error bars, and no stated match of mass, inertia, center of gravity, or drop height. The stress-test note is on target: with the evidence as presented, you cannot tell whether the similarity comes from the PBD particles or from a tuned drag/buoyancy model. The appendix says a rigid-body hydrodynamics baseline was integrated alongside PBD, but no comparison to it is shown. That comparison is exactly what the paper needs.\n\nOn the control side, the position-tracking results for hovering and the oval trajectory are reasonable for a simulator paper, though they are single-run plots without statistical backing. The RL experiments are preliminary — convergence curves for SAC and SACfD in one task — but that is acceptable for a systems/platform paper.\n\nThe central architecture is credible, and the missing evidence is obtainable. I would send this to peer review with the clear expectation of major revision: report the drop-test comparison with numbers and repeats, match the physical parameters between real and simulated AAM, show the rigid-body baseline comparison, and align the abstract with what the evaluation actually demonstrates. A reader focused on robot simulators or aerial-aquatic robots will get real value from the platform and the guidelines in the appendices.","headline":"A genuinely new integrated simulator for cross-medium manipulation, but the abstract's quantitative fidelity claim is not backed by the reported validation.","tokens_in":17185,"tokens_out":2217,"would_cite":true,"duration_ms":23004,"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":"This paper introduces a simulator, AAM-SEALS, in which an aerial-aquatic manipulator can fly, dive, and grasp objects, with particle-based water claimed to match real drop tests closely enough for simulation-first robot development.","keywords":["aerial-aquatic manipulation","cross-medium robotics","position-based dynamics","particle-based hydrodynamics","robot simulation","reinforcement learning from demonstrations","movable center of gravity control","underwater manipulation"],"falsifier":"Run repeated instrumented drops of objects with different masses, shapes, and impact speeds into a real water tank and into the simulator, and require that the acceleration traces agree quantitatively — for example, that the peak deceleration at water entry and the settling time match within a pre-specified margin across all objects. Alternatively, measure closed-loop position-tracking error during a diving trajectory on a physical prototype and compare it to the same trajectory in SEALS; if the errors diverge beyond sensor noise, the paper's claim of quantitative hydrodynamic validation would be refuted.","tokens_in":16195,"feed_emoji":"🌊","tokens_out":6675,"duration_ms":68061,"temperature":0.7,"pith_summary":"This paper introduces a new class of robot, the Aerial-Aquatic Manipulator (AAM), and a simulator, AAM-SEALS, in which such a robot can fly, swim, enter the water, and grasp objects in one continuous world. The central claim is that particle-based hydrodynamics, implemented through position-based dynamics, makes the water realistic enough to support controller design and learning without a physical robot: the authors report water-entry damping, wave impacts, buoyancy, and underwater light effects, and they validate the fluid model by dropping a 3D-printed AAM into a tank and comparing its acceleration trace to the simulation. The paper also contributes a control architecture that recomputes the rotor-force allocation matrix as the manipulator shifts the robot's center of gravity, and it demonstrates reinforcement learning, including learning from demonstrations, on tasks such as capturing a moving crab. The authors position AAM-SEALS as the first high-fidelity simulator for cross-medium manipulation and frame it as a precursor to building physical AAMs.","feed_headline":"One simulator covers flying, diving, and underwater grabbing","feed_subtitle":"Particle-based water lets aerial-aquatic robots train for sea, air, and land in a single world.","key_machinery":"The central mechanism is position-based dynamics (PBD) for the fluid: water is a discrete set of particles whose positions are iteratively corrected so that the local density satisfies C = ρ_i/ρ_0 − 1 ≤ 0, with boundary particles included in the density estimate, giving stable free-surface waves, splashes, and buoyancy at interactive rates. The second load-bearing piece is the control allocation matrix, recomputed each time step from the instantaneous center of gravity, which maps desired force and torque into rotor speeds and lets a quadrotor keep tracking a command while its manipulator moves. These two pieces are what make the claimed realism and the demonstrated cross-medium trajectories possible.","core_discovery":"The paper's central result is that a single particle-based fluid model can carry an aerial robot through the full air-to-water transition with enough realism to make simulation-first development plausible. In AAM-SEALS, water is modeled not as a rigid-body force field but as position-based dynamics (PBD) particles that satisfy a density constraint at every step, with boundary particles providing the pressures that produce buoyancy, splashes, and damping. On top of this fluid, the robot is controlled by a velocity PID and a joint PD controller, and the key mechanical idea is an allocation matrix A whose entries are recomputed continuously so the torque balance tracks the changing center of gravity as the arm moves. The evaluation claims that the resulting position-tracking error stays small in hover and on an oval trajectory that crosses the air-water boundary, and that both visual reinforcement learning and reinforcement learning from demonstrations converge in the simulator.","pith_inferences":["The paper's validation rests on a single comparison with no error metric; a careful reader should treat 'quantitatively validated' as a goal rather than a demonstrated fact until repeated trials with error statistics appear.","The dynamic center-of-gravity allocation matrix is a general idea that could transfer to any aerial manipulator with a moving payload, an implication the paper only states in the AAM context.","A natural next experiment the paper does not run is closed-loop sim-to-real transfer of a trained grasping policy; SEALS would be the right testbed for that.","The PBD fluid model may be more credible for qualitative training (splashes, damping, visual appearance) than for exact force prediction; until error metrics are reported, policies trained on it should be treated as prescreened rather than final validation."],"forward_implications":["Simulation-first development of AAMs becomes feasible: control, perception, and learning can be tested in SEALS before investing in waterproofed physical hardware.","The dynamic allocation matrix offers a template for any aerial manipulator whose payload or arm shifts the center of gravity, not only water-crossing robots.","The simulator gives robot learning a benchmark where a single policy must handle aerial, aquatic, and transitional phases, with photorealistic cameras and contact sensors for visual RL.","The search-and-capture challenge with controllable crabs and sea spiders provides a reproducible new task for evaluating cross-medium manipulation without requiring live animals in training.","If the hydrodynamic fidelity holds under broader conditions, the same PBD approach could be applied to other free-surface robotics problems, such as boats, wave energy, or flooded-environment navigation."],"supporting_citations":[{"why":"Supplies the aerial quadrotor dynamics and control structure that the paper adapts for the AAM.","marker":"[13]"},{"why":"Provides the unified position-based dynamics method used as the simulator's particle fluid engine.","marker":"[26]"},{"why":"Introduces position-based dynamics, the stability foundation for the fluid simulation.","marker":"[31]"},{"why":"Presents position-based fluids with the density constraint and boundary treatment credited for buoyancy and splashes.","marker":"[25]"},{"why":"Defines smoothed particle hydrodynamics, the underlying particle-fluid model.","marker":"[30]"},{"why":"Gives the quadrotor PID velocity control law that the paper extends to the AAM.","marker":"[29]"},{"why":"The Soft Actor-Critic algorithm used for the visual reinforcement learning evaluation.","marker":"[11]"},{"why":"Provides the soft actor-critic from demonstrations method used for the animal-search challenge.","marker":"[52]"}],"fun_headline_variants":["One particle-based world lets robots fly, dive, and grab","Simulator blends air and water for seamless robot training","Flying-swimming robot sim uses squishy particles for water","SEALS sim enables air-water robot training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the particle-based water in the simulator produces the same forces an actual robot feels when hitting and moving through water; in the paper, this is supported by only one drop-test comparison with no error metric, so if that comparison is unrepresentative, the hydrodynamic validation collapses.","fun_headline_variants_meta":{"raw":{"variants":["One particle-based world lets robots fly, dive, and grab","Simulator blends air and water for seamless robot training","Flying-swimming robot sim uses squishy particles for water","SEALS sim enables air-water robot training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000266,"raw_usage":{"total_tokens":1619,"prompt_tokens":961,"completion_tokens":658,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":577,"completion_tokens_details":{"reasoning_tokens":604}},"tokens_in":577,"tokens_out":658,"duration_ms":7289,"temperature":1.0,"reasoning_tokens":604,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:53:32.046166+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run repeated instrumented drops of objects with different masses, shapes, and impact speeds into a real water tank and into the simulator, and require that the acceleration traces agree quantitatively — for example, that the peak deceleration at water entry and the settling time match within a pre-specified margin across all objects. Alternatively, measure closed-loop position-tracking error during a diving trajectory on a physical prototype and compare it to the same trajectory in SEALS; if the errors diverge beyond sensor noise, the paper's claim of quantitative hydrodynamic validation would be refuted.","supporting_citations":[{"cited_title":"Pegasus Simulator: An Isaac Sim Framework for Multiple Aerial Vehicles Simulation","cited_arxiv_id":"2307.05263","evidence_quote":"Supplies the aerial quadrotor dynamics and control structure that the paper adapts for the AAM."},{"cited_title":"Unified particle physics for real-time applications","cited_arxiv_id":null,"evidence_quote":"Provides the unified position-based dynamics method used as the simulator's particle fluid engine."},{"cited_title":"Position based dynamics","cited_arxiv_id":null,"evidence_quote":"Introduces position-based dynamics, the stability foundation for the fluid simulation."},{"cited_title":"Position based fluids","cited_arxiv_id":null,"evidence_quote":"Presents position-based fluids with the density constraint and boundary treatment credited for buoyancy and splashes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines smoothed particle hydrodynamics, the underlying particle-fluid model."},{"cited_title":"Learning from ambiguous demonstrations with self- explanation guided reinforcement learning","cited_arxiv_id":null,"evidence_quote":"Provides the soft actor-critic from demonstrations method used for the animal-search challenge."}],"review_version":1}