acceptodds
Under review as a conference paper at ICLR 2027

Energy-Efficient Motion Control in Fluid-Structure interaction

Abstract

Energy-efficient motion through a fluid requires coordinating body movement with fluid forces and torques. This task is challenging because body motion changes the surrounding flow, which in turn affects how the body moves. Here, we study body motion control in flowing channels using a training dataset generated with computational fluid dynamics (CFD). The dataset covers two body shapes, three channel geometries, and both sensor-free and sensor-based settings. We consider direct trajectory control and force-and-torque control within a common framework that enables comparisons of physical predictors and control generators across different settings. For sensor-free circular-body control, experiments identify paths using up to 25.7% less work than the corresponding best training trajectories and demonstrate feasible planning with unseen endpoints. For sensor-based square-body control, the best neural controllers use 36.6% and 64.4% less work than the corresponding best training trajectories in the unobstructed and central-obstacle channels, respectively, and 32.3% less in the more complex two-wall channel. We hope this work will help extend the use of physical operators beyond forward prediction to more challenging control tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
fluid–structure interaction, neural operator, computational fluid dynamics, partial differential equation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_rJTJOpenReview ↗
Rating
3(+2)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies rigid-body motion control in quasi-2D CFD simulations at Re = 100. It combines established physical predictors with various optimization and control algorithms and reports work reductions relative to training demonstrations. A hybrid controller achieves high completion rates.

Critical strengths
  1. Extensive implementation and evaluation of multiple predictors and control algorithms.
  2. Controls are evaluated in CFD, and unsuccessful configurations are reported rather than discarded.
Critical weaknesses
  1. A major concern is about what the paper is solving rather than how. Neither experimental setup corresponds to a physical control problem. The circular body follows an externally prescribed trajectory at constant speed, so its motion is not controlled by anything in the formulation. The square body is driven by directly imposed forces and torques with no propulsion or actuation mechanism behind them. Neither is a stylized version of a real navigation or fluid–structure system and they are missing precisely the features that make such systems difficult. The setting compounds this. Quasi-2D laminar flow at Re=100Re = 100, fixed geometries, fixed body properties, and essentially fixed tasks. The experiments do not vary flow regime, actuation dynamics, or observability, yet these are exactly the ingredients of the applications the introduction gestures toward. Nowhere does the paper state which essential difficulty of real navigation its formulation is designed to isolate and preserve. An idealized benchmark needs no immediate applicability but it has to expose a transferable scientific principle, and after working through the paper I cannot identify one.

  2. Beyond the significance concern above, I am unable to point to a contribution that is new. The predictors are established architectures (FNO, U-Net, Transolver, POD-Kalman), the controllers are established methods (CEM, MPPI, PPO, DOB), and the paper's own framing is the combinations among them. Combining known components can of course be a contribution in itself but then the combinations have to be the object of study, and the study has to yield a finding about them. As far as I can tell, this one does not. The experimental section, extensive as it is, stops short of exactly that. For instance, POD-Kalman posts among the worst field-prediction errors (Fig. S13) and yet achieves the lowest control cost in the two-wall setting, beating every neural predictor-controller combination. If a three-decades-old reduced-order model with poor prediction accuracy outperforms all the learned surrogates, the central motivation, that accurate neural flow prediction improves downstream control, fails to hold in the paper's own experiments. This is either an important negative finding or a symptom of an ill-posed evaluation. What remains is a large performance matrix without an organizing question. The paper does not establish what determines control quality, when a learned model helps, or any condition under which one predictor or controller should be preferred. Without such a finding, the study demonstrates that the combinations can be run, not that they teach us anything.

  3. In my opinion, some of the observations are problematic. In particular, (i) DOB alone without learning reduces work from 26.104 J to 14.721 J around the central obstacle (Table 1), a 43.6% improvement. Thus, most of the headline 64.4% reduction is already achieved without a learned predictor. (ii) the conventional reduced-order model POD-Kalman outperforms all neural controllers in the two-wall setting, (iii) In Fig. 6e, successful central-obstacle hybrid configurations generally achieve approximately 56–64% savings, largely independently of predictor architecture, optimizer, or sensor layout.

  4. The main method comparison appears confounded. Hybrid planning achieves 382/396 completions compared with 51/396 for four-second predictor-based planning. However, Appendix F shows that these approaches differ in their correction bounds, objective functions, action mappings, acceptance rules, and use of additional forecasts. This is not a controlled comparison establishing the benefit of hybrid prediction. The reported improvement could arise from the substantially different controller design. There is no rigid-body-only MPC baseline with identical optimization and acceptance mechanisms. Consequently, even the paper's strongest result does not demonstrate an advantage from learned fluid dynamics.

  5. The circular body's baseline is the best of its training trajectories, and the square body follows feedback controllers with prescribed rotation modes denoting prescribed references, not optimized ones. Outperforming replayed or scripted behavior is weak evidence of effective optimization and the natural comparison of the same tasks under an optimized classical controller is absent. Additionally, the computational account is missing precisely where it matters. The hybrid scheme allows 8192 model-transition equivalents per planning decision, which is substantial, yet runtime and scaling are not reported anywhere. For a method whose selling point is planning quality, one needs at least a cost-benefit comparison against a conventional controller of comparable budget.

  6. The energy objective is not well justified. Eq. 2 takes the positive part of combined translational and rotational actuator power. Negative rotational work can therefore cancel positive translational work. This assumes ideal energy transfer or recovery between distinct actuation channels, without any corresponding physical mechanism. Given that energy efficiency is the paper's central objective, this modeling choice directly affects the interpretation of the claimed improvements.

Questions for authors
  1. What concrete physical system corresponds to the assumed actuation and flow conditions? What scientifically important difficulty does this idealized setup isolate?
  2. What specific result established by this paper would remain relevant beyond the three prescribed CFD configurations? Which existing understanding of learning-based physical control does it advance?
  3. Given the DOB results in Table 1, POD-Kalman's superior two-wall performance, and the near-uniform hybrid results in Fig. 6e, what evidence establishes that neural fluid prediction is beneficial?
  4. How can the 382/396 versus 51/396 comparison support a conclusion about hybrid prediction when the two approaches use different action constraints, objectives, and acceptance rules?
6 comments

Weakness 1 / Question 1. What physical systems do the two tasks model, and what scientific difficulty do they retain?

The circular task models towing a body through a confined current at 0.20 m/s. The planner chooses the route. The flow along that route determines the opposing load and towing work. The square models planar motion under a commanded net force and torque. The maximum applied force magnitude is 100 N, and the maximum applied torque magnitude is 1.2 N·m. The planner chooses their time history so the body clears the obstacles, reaches the target and stops. The benchmark uses 2D flow at Re = 100, and the feedback tests use 0, 8, 48 and 120 pressure channels (Section 2.1 and Appendices B.1 and G.3).

Much of PDE benchmarking focuses on forward problems under prescribed conditions, and many flow benchmarks use fixed bodies. We formulate and benchmark an inverse control problem with two-way fluid–structure coupling. The target and motion constraints are given, and the planner must find a path or an actuation sequence that completes the task with low work.

In the square task, each candidate force and torque sequence changes how the body translates and rotates. This motion changes the fluid boundary and the subsequent flow. The evolving flow then changes the loads on the body and affects its later motion. The search is more demanding than path planning in a prescribed environment because a candidate's feasibility and work depend on the flow it creates and the body's response. We aim to advance research on inverse control in coupled physical systems and examine how learned forward models can guide that search (Sections 1, 2.1 and 5).

Weakness 2 / Question 2. What finding advances the evaluation of learned physical predictors for control?

Field error measures prediction on a prescribed test trajectory. In control, the forecast helps choose the trajectory, and the work depends on the loads and motion along that chosen trajectory. We compare these rankings within groups with the same geometry, pressure layout and search method. The six neural field predictors and POD–Kalman supply comparable field outputs. Among successful controllers, 44 of 48 matched groups have a negative Spearman correlation between held-out field MSE and executed work. The median is −0.60 overall and −0.10 between the walls. The median remains −0.60 when the analysis uses the six neural field predictors (Fig. 8c and Appendix G.7).

The result shows why the evaluation of a control predictor needs to measure the feasibility and work of the motion it selects in coupled CFD. Field error and realized work give different rankings, and the strength of the association varies with geometry (Sections 4–5).

Weakness 3 / Question 3. How much of the saving is already achieved by DOB, and what do the full distribution and POD–Kalman comparison show?

At the central obstacle, DOB reduces work by 43.7% relative to the 26.200 J training reference. The best hybrid run uses 38.9% less work than DOB. Across the 251 successful hybrid configurations, the middle half of the work values lies between 9.913 and 10.586 J. Four hybrid configurations use less work than plain PD (Section 3.2, Table 1 and Appendices D.4 and G.4).

GeometryPD (J)DOB (J)Best hybrid (J)Median hybrid (J)
No obstacle7.88614.0027.6998.892
Central obstacle9.16914.7499.00910.271
Two wallsFailed49.58536.26541.291

In the revised two-wall sweep, POD–Kalman and Load Transformer give comparable minimum work of 36.313 and 36.265 J, respectively (Fig. 6e and Table 1). The paired forecast comparison fixes the learned proposal generator and DOB and varies the forecast used for scoring and feasibility. With both forecasts, work falls in 219 of the 272 jointly successful pairs, with a median saving of 7.9% (Fig. 8b, Table 3 and Appendix G.2). Feedback provides a substantial share of the saving relative to the training reference. The paired comparison measures the additional work saving from learned forecasting on top of that feedback.

Weakness 4 / Question 4. What can be attributed to the score, the full controller design and the learned forecast?

The original 382/396 versus 51/396 comparison evaluates two complete planners. We swap Eq. (11) for Eq. (7) within single-forecast planning, keeping the checkpoint, search, action map, feasibility checks and acceptance rule fixed. The revised score raises completion from 54/528 to 67/528 at 1 s and from 56/528 to 69/528 at 4 s. Completion totals are similar at both horizons under either score. The full hybrid design completes 489/528 on the same grid (Table 2 and Appendix G.1).

The full planner bounds the corrections to DOB, tapers them near the target and uses progress, feasibility and score-improvement checks before acceptance. With this design and the learned proposal generator shared, hybrid and constant-load scoring complete 299/336 and 301/336. Hybrid scoring lowers work in 219 of 272 joint successes, with median savings of 11.0%, 10.5% and 3.3% across the three geometries and 7.9% overall. The full design achieves the larger completion gain. The controlled forecast comparison finds similar completion totals and a work saving among joint successes (Section 5, Fig. 8a–b, Table 3 and Appendices F.3 and G.2).

Weakness 5a. How were the training references generated, and which classical controllers are evaluated on the same tasks?

The circular references are the lowest-work paths from the corresponding training groups. The paths were generated by sampling control points, smoothing and sampling at the prescribed speed. Square demonstrations use feedback with prescribed rotation modes. In control evaluation the terminal angle is measured modulo the square's rotational symmetry. We evaluate PD, DOB and optimized constant-load MPPI on the same square tasks, using the training trajectories as a separate reference (Appendices B.2, D.1–D.2 and G.4).

Weakness 5b. What are the work and runtime at comparable planning budgets?

Table 5 compares the deadline-aware planners under a cap of 8,192 model transitions per decision with shared terminal and progress checks and residual bounds. Candidate rollouts use the relevant forecasts. OpenFOAM executes and scores the selected controls. Work is separate positive work, reported as a median over successful runs in the three pressure layouts (Appendix G.4).

PlannerCompleteNo obstacle (J)Central obstacle (J)Two walls (J)
Deadline DOB6/99.2979.244—
Constant-load MPPI9/99.77611.48242.504
Transolver + guard9/98.10310.08540.463
RSSM + guard9/98.12010.19944.441
RSSM + guard, Dreamer proposal9/97.8549.91044.637

Constant-load MPPI completes all three tasks. Learned variants use less work in 22/27 matched scene–layout comparisons. Deadline DOB uses the least work at the central obstacle among the planners in Table 5.

Table 8 compares ordinary MPPI with the same number of candidates, using Load-TCN in the learned branch. It uses 4,096 model transitions per decision at the full horizon with one forecast and 8,192 with both. All six runs complete. Work is separate positive work on the executed trajectories. Planning times are averages over 108 decisions replayed per run on the same A100 (Appendix G.8).

GeometryRigid-body onlyLearned onlyHybrid
Central obstacle11.650 J / 0.79 s10.164 J / 0.90 s10.284 J / 1.70 s
Two walls44.361 J / 0.81 s38.203 J / 0.91 s41.131 J / 1.71 s

In these matched tasks, learned-only uses the least work and rigid-body-only has the shortest planning time. The two-forecast planner takes about twice the per-decision time of either single-forecast variant at equal candidate counts.

Weakness 6. How does the revised objective handle braking, and how does the change affect completion and work?

The original objective takes the positive part after adding translational and rotational power. A negative contribution from one component can offset a positive contribution from the other. The revised Eq. (2) takes the positive part of each component before adding them. Braking contributes zero in its own component, and positive work in the other component is counted in full. The revised square-body sweep uses this separate-positive-work objective (Sections 2.1, 3.2–3.3 and Appendix G.5).

We rerun hybrid planning under both work objectives using the finer recorded CFD time points. The predictors and applicable policies are trained separately under each objective with the same training protocol. Each objective covers three geometries, two load predictors, seven search or policy-guided methods and four pressure layouts. Both sets of trajectories are evaluated with the same separate-positive-work metric (Table 7 and Appendix G.6).

GeometryCombined objective completeSeparate objective completeBoth completeMedian paired work change
No obstacle53/5652/5649−0.68%
Central obstacle50/5649/5643−1.76%
Two walls45/5648/5639−0.81%
All148/168149/168131−1.29%

Overall completion is similar under the two objectives. In 81 of 131 joint successes, the separate objective uses less work. The interquartile range of the paired change is −5.18% to +1.17%, and the median is below zero in all three geometries.

Reviewer_YY8LOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a benchmark for evaluating combinations of a flow predictor and a controller for a body moving through a channel, with the goal of reaching a target with minimum work. For flow prediction, several architectures are compared: FNO, F-FNO, UNO, U-Net, Transolver, and Swin, together with load predictors (Load-GRU/TCN/Transformer), path-based work predictors (PathCNN, Path Transformer), POD-Kalman, and a constant-load rigid-body model. For control generation, the paper compares CEM, gradient search, PPO, SAC, TD3, Dreamer, shooting, and MPPI, and additionally evaluates imitation- and offline-RL-based direct controllers (BC, DT, diffusion policy, IQL, CQL) as baselines. Combinations are evaluated on task completion, positive work relative to a training reference, and failure-mode classification, across three channel geometries, two sensor-free/sensor-based settings, several sensor layouts, and 1 s vs. 4 s planning horizons. The body is a 2D circle or a 2D square.

Critical strengths

The paper provides a shared framework in which predictors and control generators can be compared under identical task conditions, and it evaluates these combinations across a wide range of settings. Reporting the fully negative sensor-free square result is useful. The hybrid planning results (learned predictor combined with a constant-load rigid-body model) are a practically relevant observation, since extending the planning horizon alone improves completion only marginally (8.8% to 12.9%), whereas hybrid planning reaches roughly 95%.

Critical weaknesses

Limited methodological novelty. The main contribution is a benchmark and an empirical grid evaluation rather than a new algorithm. The hybrid planning scheme is essentially a learned model combined with a simple fallback model inside a receding-horizon loop, which is close to existing robust MPC ideas.

Selection and statistical concerns. Best combinations are reported after searching a large configuration space, so the headline reductions (25.7%, 36.6%, 64.4%, 32.3%) may partly reflect selection over many trials. Seed variance, confidence intervals, and statistical significance are not reported. Comparisons are made against a per-model "best training reference," which differs across models; a common, strong classical baseline (e.g., CFD-in-the-loop nonlinear MPC) would make the claims more convincing.

Insufficient ablation of hybrid planning. It is unclear how much of the completion improvement comes from the learned predictor, from the rigid-body fallback acting as a safety filter, or from the base DOB feedback. A decomposition is needed to support the claim that hybrid planning is the key ingredient.

Limited scope and generality. Evaluation is restricted to 2D, Re = 100, two shapes, and three channel geometries, all in simulation. How the benchmark and conclusions extend to 3D, higher Reynolds numbers, and real-world conditions is not discussed. As a benchmark, this limits its generality.

Reproducibility. Code and data are stated to be released only after review, which weakens the benchmark's immediate value to the community.

Questions for authors

The reported results could be of interest to readers working on learned surrogates for flow control, and the shared evaluation protocol has some value. However, the methodological novelty is limited and the analysis is largely a thorough grid evaluation of existing predictors and controllers, without ablations or statistical support for the main claims. The scope is narrow, and much of the reported comparison would plausibly be reproduced by practitioners for their own specific setup. I am not convinced the benchmark will be broadly adopted by the community, and I lean toward rejection in its current form. A revised version that adds ablation of the hybrid scheme, statistical reporting, a common classical baseline, and a discussion of generality would substantially strengthen the case.

5 comments

Weakness 1. What is the scientific contribution beyond combining existing predictors and controllers?

PDE benchmarks have largely focused on forward problems under prescribed conditions, and many flow benchmarks keep the body fixed. Our contribution is to formulate and benchmark the inverse problem of finding energy-efficient motion under two-way fluid–structure coupling. Given a target and motion constraints, we seek a path or a sequence of applied forces and torques that completes the task with low work.

A path search in a prescribed environment evaluates routes against given conditions. In our problem, the body's translation and rotation change the fluid boundary and alter how the flow evolves. The evolving flow in turn changes the forces and torques on the body and affects its subsequent motion. This two-way coupling makes the optimization more demanding because the planner must account for the coupled evolution of the body and the flow under each candidate control.

Solving inverse problems is essential to using physical models to achieve desired outcomes. We aim to advance research on inverse problems in coupled physical systems by providing a benchmark for studying how learned forward models can guide the search for feasible, low-work controls (Sections 1, 2.1 and 5).

Weakness 2a. How were headline results selected, and what variation do the statistics describe?

For preplanned controls, each predictor–search combination can submit three candidates per geometry. The method nominates all three before CFD evaluation, and every submission counts. This gives 216 circular paths per endpoint setting and 243 preplanned square controls (Appendix B.4).

For closed-loop square control, the four-layout hybrid sweep completes 788/864 configurations. The three nonzero-pressure layouts in Fig. 6e contribute 593/648, and Fig. 6b shows their minimum-work motions. Across the full sweep, the completion counts are 288/288, 251/288 and 249/288 by geometry. All 788 successful runs use less work than the same-task DOB. The separate forecast comparison gives a median saving of 7.9% across 272 joint successes, with an interquartile range of 1.4%–13.4%. Each configuration uses one evaluation seed, so the distributions describe variation across configurations (Fig. 8b, Table 3 and Appendices D.4 and G.2–G.3).

Weakness 2b. Are the training references shared, and how does hybrid planning compare with same-task PD and DOB?

The main square-body comparison uses one training reference per geometry. The circular task shares references within the field-predictor and path-predictor groups. The work below is in joules, and hybrid medians and quartiles cover the four pressure layouts (Section 3.2, Table 1 and Appendices B.4, D.4 and G.4).

GeometryTraining referencePDDOBBest hybridHybrid median [Q1, Q3]
No obstacle12.1127.88614.0027.6998.892 [8.250, 9.520]
Central obstacle26.2009.16914.7499.00910.271 [9.913, 10.586]
Two walls53.587Failed49.58536.26541.291 [39.893, 43.005]

Hybrid median work is 16.7%–36.5% below same-task DOB. PD uses less work than the hybrid median in the two simpler geometries. Between the walls, hybrid planning completes the task and uses less work than DOB.

Weakness 2c. How does a classical optimizer perform under a common search budget?

We compare deadline-aware planners with the same terminal and progress checks and a cap of 8,192 model transitions per decision. Classical MPPI searches with the constant-load rigid-body forecast, and OpenFOAM executes and evaluates the resulting controls. The work values are medians of separate positive work over completed runs in the three pressure layouts (Table 5 and Appendix G.4).

PlannerCompleteNo obstacle (J)Central obstacle (J)Two walls (J)
Deadline DOB6/99.2979.244—
Constant-load MPPI9/99.77611.48242.504
Transolver + guard9/98.10310.08540.463
RSSM + guard9/98.12010.19944.441
RSSM + guard, Dreamer proposal9/97.8549.91044.637

The pressure-independent classical planner provides the same motion for a given geometry across the three layouts. Constant-load MPPI completes all three tasks. The learned variants improve on its work in 22/27 matched scene–layout comparisons. Deadline DOB has the lowest work at the central obstacle among these planners.

Weakness 3a. What accounts for the completion gain in the hybrid planner?

We compare the original cost in Eq. (11) with the learned-forecast score in Eq. (7), holding the checkpoints, search, action map, feasibility checks and acceptance rule fixed. Completion rises from 54/528 to 67/528 at 1 s and from 56/528 to 69/528 at 4 s. The two horizon settings have similar completion totals under either score. The full hybrid planner completes 489/528 on the same four-layout grid (Table 2 and Appendix G.1).

The full design applies corrections to the DOB action, limits their size and tapers them near the target. A correction must pass relative-progress and feasibility checks and improve the score by the acceptance margin. The controller returns the DOB action when a correction is rejected. DOB alone completes all three tasks with work of 14.002, 14.749 and 49.585 J by geometry. Changing the score improves completion within single-forecast planning, and the much larger completion gain is observed for the full DOB-based design (Sections 3.2 and 5, Table 1 and Appendices F.2–F.3 and G.4).

Weakness 3b. What does the learned forecast contribute, and how does it interact with the rigid-body forecast?

The learned forecast updates fluid loads along each candidate motion. The rigid-body forecast holds the current load estimate fixed. The hybrid planner averages their scores and checks feasibility and progress under both forecasts. In the matched policy-guided MPPI comparison, the learned proposal generator, DOB feedback, residual map, check rules and candidate count are shared. Hybrid and rigid-body scoring complete 299/336 and 301/336 configurations. Hybrid uses less work in 219 of 272 joint successes, with a median saving of 7.9%. In this matched comparison, the added learned forecast mainly reduces work among feasible motions (Fig. 8a–b, Table 3 and Appendix G.2).

We also compare all three forecast sets with ordinary MPPI at the same candidate count. All six runs complete. The learned branch uses Load-TCN. Work is measured as separate positive work on the executed trajectories, and the times are means of 108 decisions replayed per run on the same A100 (Table 8 and Appendix G.8).

GeometryRigid-body onlyLearned onlyHybrid
Central obstacle11.650 J / 0.79 s10.164 J / 0.90 s10.284 J / 1.70 s
Two walls44.361 J / 0.81 s38.203 J / 0.91 s41.131 J / 1.71 s

Learned-only gives the lowest work in both tasks. The two-forecast planner evaluates twice as many model transitions at matched candidate counts, and its mean planning time is about twice that of either single-forecast variant in these tests.

Weakness 4. What transfer was tested, and how would the benchmark extend to new flow regimes or physical systems?

The circular transfer experiment shifts both start and target and replans with the trained predictors fixed. All 216 paths are feasible across the three geometries. Their minimum work values are 3.171, 3.193 and 5.466 J in scene order (Section 3.1, Fig. 5 and Appendix D.1).

The present benchmark uses two shapes in quasi-2D laminar flow at Re = 100. The same inverse-control formulation extends to 3D by coupling six-degree-of-freedom body dynamics to the flow. Evaluation at higher Reynolds numbers calls for data and a solver appropriate to the target flow regime. A physical benchmark would use measured actuator and sensor dynamics to evaluate the chosen controls, task constraints and work (Section 2.1 and Appendix B.1).

Weakness 5. When will the code and dataset be public?

We plan to release the code and dataset publicly after peer review. The raw training and validation data total about 2 TB. We are identifying a platform to host the data.

Area_Chair_KaKDOpenReview ↗

Feedback

Hi, the review deadline is approaching. Please take a look at the authors' rebuttal and post your feedback (and update your rating if appropriate) at your earliest convenience. Thank you! AC

Reviewer_Vs5XOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies energy-efficient control of bodies moving through fluid-filled channels. It introduces CFD-generated trajectories for circular and square bodies in three channel geometries. Circular-body control optimizes a prescribed constant-speed path, while square-body control optimizes applied forces and torque subject to collision avoidance and terminal stopping constraints.

The study compares numerous learned predictors and control generators in open-loop, observation-based, and hybrid planning settings. The hybrid controller combines disturbance-observer feedback with corrections evaluated using both a learned predictor and a rigid-body model with constant estimated fluid loads. Selected controls are evaluated in OpenFOAM. The results show work reductions relative to training demonstrations and a DOB baseline, substantially higher completion for the complete hybrid configuration, and failure of all evaluated sensor-free square-body controls.

Critical strengths

Evaluation measures executed control outcomes rather than surrogate predictions alone. Candidate controls are tested in OpenFOAM against collision, actuator, and terminal-state constraints. This exposes substantial failures that prediction metrics alone would miss: none of the 243 selected sensor-free square-body controls completes the task. Reporting these negative results makes the study more useful for understanding the difficulty of applying learned physical models to control.

The selected successful controllers achieve meaningful work reductions against a feedback baseline. Appendix D.2 reports work reductions of 45.2%, 36.9%, and 27.2% relative to DOB in the unobstructed, central-obstacle, and two-wall channels, respectively. These comparisons provide stronger evidence than improvements over the best training demonstration alone. The range of predictor families, including direct-load models and POD–Kalman, also provides a useful starting point for a comparative benchmark.

Critical weaknesses

The central hybrid-planning comparison does not isolate the contribution of the learned predictor or the additional dynamics model. Completion increases from 51/396 configurations with four-second predictor-based planning to 382/396 with hybrid planning. However, Appendices F.2–F.3 show that the hybrid configuration also introduces a different objective, explicit deadline and stopping penalties, relative-progress checks, and an acceptance margin. Reusing the same learned checkpoints therefore does not make this a controlled test of the second forecast. A particularly important missing baseline is constant-load rigid-body MPC using the same DOB controller, action parameterization, objective, safeguards, and search budget, but no learned predictor. Without this comparison, it remains unclear whether the learned physical models are necessary for the strongest results.

The evaluation breadth is primarily across algorithm configurations, with limited evidence of controller robustness or generalization. The headline 615/648 completion figure counts different predictor–generator–geometry–sensor combinations, rather than repeated independent trials of a fixed controller. The highlighted work reductions are selected minima across a large search over configurations. Square-body control uses prescribed endpoints and three fixed geometries; the explicit endpoint-transfer experiment concerns the easier prescribed-path circular task. Held-out prediction trajectories do not establish control generalization. Repeated runs of representative controllers over varied initial conditions, flow conditions, or goals are needed to support reliable performance comparisons and broader benchmark conclusions.

Questions for authors

Component attribution: How does constant-load rigid-body MPC perform when it uses exactly the hybrid controller’s objective, action mapping, progress checks, stopping penalties, and search budget, but omits the learned forecast? Conversely, how does the learned-only planner perform with these same safeguards? This would distinguish model benefits from controller-design changes.

Evaluation and variability: For representative fixed controllers, what are the completion rates and work distributions across independent training/search seeds and held-out variations in initial state, inflow, or target? Please distinguish these statistics from completion counts over heterogeneous algorithm configurations.

Value of pressure sensing: What happens when the pressure observations are removed but body-state and action histories are retained? The disturbance observer in Equation (8) already estimates fluid loads from velocities and known actions. A state-only baseline would establish whether the pressure sensors provide additional control value.

Numerical validation of work savings: Equation (10) estimates positive work using interval-endpoint samples at 0.25 s spacing. Do the rankings and reported savings persist when work is integrated at CFD substeps and the CFD mesh/time step is refined? Please also clarify that clipping the sum of translational and rotational power assumes that negative power in one component can offset positive power in another; how sensitive are the results to summing their positive parts separately?

3 comments

Weakness 1 / Question 1a. How much of the original 51/396 versus 382/396 completion gap comes from changing the score?

The original counts compare complete planning designs. To measure the effect of the score, we replace the original cost in Eq. (11) with the learned-forecast score in Eq. (7) within the single-forecast planner. The checkpoint, search, action map, feasibility checks and acceptance rule are held fixed. The revised score evaluates work and progress relative to zero correction and includes deadline and stopping penalties (Appendix G.1).

HorizonOriginal scoreRevised score
1 s54/52867/528
4 s56/52869/528

The grid covers 11 predictors, four search methods, three geometries and four pressure layouts. The score change adds 13 completions at each horizon. Under either score, 1 and 4 s give similar completion totals. Among the 29 pairs that complete under both scores at 4 s, the revised score lowers work in 28, with a median reduction of 26.6% (Table 2 and Appendix G.1).

The full hybrid planner completes 489/528 configurations on the same grid. It applies bounded corrections to the DOB action, reduces their amplitude near the target, and uses progress, feasibility and score-improvement checks to select corrections. The larger completion gain is observed for the full DOB-based planning design (Section 5 and Appendices F.2–F.3).

Question 1b. How do constant-load-only and learned-only planning compare with hybrid when its safeguards are shared?

We compare three forecast sets in ordinary MPPI at the central obstacle and between the walls. The controller shares the four-second maximum horizon, DOB action, residual map, objective terms, acceptance rule and candidate counts across variants. The progress and feasibility checks use the active forecast set. The learned branch uses Load-TCN. Each single forecast uses 4,096 model transitions at the full horizon, and the hybrid uses 8,192 at the same candidate count. All six runs complete (Table 8 and Appendix G.8).

GeometryRigid-body onlyLearned onlyHybrid
Central obstacle11.650 J10.164 J10.284 J
Two walls44.361 J38.203 J41.131 J

The reported values use separate positive work on the executed trajectories. Learned-only uses the least work in both tasks. In the complementary comparison under a common cap of 8,192 transitions per decision, constant-load MPPI completes all three tasks and the learned variants use less work in 22/27 matched scene–layout comparisons (Table 5 and Appendix G.4).

Question 1c. What does adding the learned forecast contribute when proposal generation and feedback are fixed?

We match 336 policy-guided MPPI configurations across three scenes and four pressure layouts. The same learned proposal generator, DOB controller, residual map, objective terms, checks and candidate count are used in each pair. The generator uses each run's observation history. The rigid-body arm scores candidates and checks feasibility with the constant-load forecast. The hybrid arm uses both forecasts (Appendix G.2).

GeometryHybrid completeRB completeBoth completeHybrid uses less workMedian saving
No obstacle112/112112/1121128411.0%
Central obstacle95/112102/112868110.5%
Two walls92/11287/11274543.3%
All299/336301/3362722197.9%

Completion totals are close, with different sets of completed configurations in the two obstacle geometries. The learned forecast chiefly contributes lower work among joint successes. The middle half of the paired savings lies between 1.4% and 13.4% (Fig. 8a–b and Table 3).

Weakness 2 / Question 2. How were candidate runs selected, what do the distributions measure, and what transfers to held-out tasks?

For preplanned control, each predictor–search combination selects three candidates per geometry before CFD execution. All three submissions count toward completion. For closed-loop square control, each configuration is evaluated with one seed at its prescribed start, inflow and target. The four-layout hybrid sweep completes 788/864 configurations, including 593/648 in the three nonzero-pressure layouts plotted in Fig. 6e. Every completed configuration uses less work than the same-task DOB. Fig. 6b shows the minimum-work trajectories, and the successful runs give the following medians and quartiles (Appendices B.4, D.4 and G.2–G.3).

GeometryCompleteMedian work [Q1, Q3] (J)
No obstacle288/2888.892 [8.250, 9.520]
Central obstacle251/28810.271 [9.913, 10.586]
Two walls249/28841.291 [39.893, 43.005]

These counts and intervals measure variation across configurations. The control-transfer test shifts both circular endpoints and replans with the trained predictors fixed. All 216 replanned paths satisfy the task constraints, with minimum work of 3.171, 3.193 and 5.466 J by geometry (Section 3.1, Fig. 5 and Appendix D.1).

Question 3. What changes when pressure is removed and body-state and action histories are retained?

Masking pressure leaves body-state and past-action histories available to the controller and keeps DOB active. With pressure masked, 195/216 configurations complete across the three geometries. Table 4 compares all four pressure layouts. The paired work change uses configurations that complete with both zero and eight channels (Appendix G.3).

Geometry0 channels8 channels48 channels120 channelsMedian paired work change, 0 to 8
No obstacle72/7272/7272/7272/72−0.03%
Central obstacle60/7265/7261/7265/72−1.00%
Two walls63/7260/7264/7262/72+1.80%

Eight channels improve completion and reduce median paired work at the central obstacle. Between the walls, four-second field MSE for the six neural field models falls from 0.435 to 0.379, while completion falls and paired work increases. Pressure can improve field accuracy, and its control effect depends on the geometry.

We also test whether feedback can generate viable fixed action sequences. Preplanning the full action sequence yields 0/243 completions. Feedback generates 161 sequences in 243 attempts, and 157 of those 161 complete when replayed as fixed actions from the original initial state (Section 3.3 and Appendix G.3).

Question 4. Do the work rankings and savings persist at finer CFD time points and under separate clipping?

We rescore the trajectories at finer recorded samples. For 623 successful combined-work runs, the work rankings within each geometry have Kendall's τ of 0.993, 0.991 and 0.969. The median work changes from using the finer samples are −1.03%, −1.07% and −2.79%, respectively. Thus the rankings largely persist under finer work sampling (Appendix G.5).

In Eq. (2), translation and rotation contribute their positive power separately. Braking in either component contributes zero in that component, and positive power in the other is counted in full. At the finer points, separate clipping raises median measured work by 0.17%, 0.17% and 0.70% by geometry. For the three archived neural controls, the original savings against DOB of 45.2%, 36.9% and 27.2% become 45.4%, 37.0% and 27.2% after rescoring at finer points with separate positive work (Appendix G.5).

We also rerun planning with each work objective at the finer CFD time points, covering three scenes, two load predictors, seven search or policy-guided methods and four pressure layouts. The separate and combined objectives complete 149/168 and 148/168 configurations. Under the same separate-positive-work measure, the separate objective lowers work in 81 of 131 joint successes, with a median saving of 1.29%. The two objectives give similar total completion, with lower median paired work under separate clipping (Table 7 and Appendix G.6).

Area_Chair_KaKDOpenReview ↗

Feedback

Hi, the review deadline is approaching. Please take a look at the authors' rebuttal and post your feedback (and update your rating if appropriate) at your earliest convenience. Thank you! AC

Comments on the paper

2 comments

Dear Reviewers,

Thank you for your feedback. The benchmark asks how to use forward physical prediction to solve an inverse control problem. Given a target and motion constraints, the planner chooses the path or the applied force and torque. For the force-and-torque task, translation and rotation change the fluid boundary and the evolving flow, which in turn changes the forces and torques on the body. We evaluate the resulting motions in OpenFOAM by task completion and actuator work (Sections 2.1 and 5).

  • Square-body work and main results. Equation (2) now counts positive translational and rotational work separately. We retrained the square-body predictors and applicable policies and reran the main experiments. The updated hybrid sweep completes 788/864 configurations across three geometries and four pressure layouts. All 788 successful runs use less work than same-task DOB. Fig. 7a now reports time-resolved velocity errors on 30 held-out trajectories using the retrained predictors (Sections 3.2–3.3, Figs. 3, 6–7 and Appendix E).

  • Planner score and forecast ablations. Replacing the original single-forecast cost in Eq. (11) with Eq. (7) adds 13 completions out of 528 at each horizon. Completion totals are similar at 1 and 4 s under either score. The full hybrid design completes 489/528 on the same configuration grid. With the learned proposal generator and DOB feedback fixed, hybrid and rigid-body scoring complete 299/336 and 301/336. Hybrid reduces work in 219 of 272 jointly successful pairs, with a median saving of 7.9% (Sections 4–5, Fig. 8a–b, Tables 2–3 and Appendices G.1–G.2).

  • Classical baselines and planning cost. We report same-task PD and DOB results. Under a common transition cap, constant-load MPPI and the learned variants complete all three tasks. On the recorded trajectories, constant-load MPPI uses 9.776, 11.482 and 42.504 J of separate positive work in scene order, and the learned variants use less work in 22/27 matched comparisons. With candidate counts fixed and Load-TCN in the learned branch, rigid-body-only, learned-only and hybrid ordinary MPPI average 0.79–0.81, 0.90–0.91 and 1.70–1.71 s per decision, respectively, on the same A100. Learned-only uses the least separate positive work in both timed tasks, at 10.164 and 38.203 J (Tables 5 and 8, Appendices G.4 and G.8).

  • Pressure observations and feedback. With pressure masked, state and action history plus DOB support 195/216 completions. The four pressure layouts give 195–199 completions out of 216 each, with geometry-dependent effects on paired work. Of 243 feedback-generation attempts, 161 produce fixed sequences for replay from the original initial state, and 157/161 replays complete. Preplanning the full sequence gives 0/243 completions (Section 3.3, Table 4 and Appendix G.3).

  • Work sensitivity and prediction quality. Rescoring 623 successful trajectories at finer recorded time points largely preserves the work rankings. In the matched objective reruns, separate and combined positive work give 149/168 and 148/168 completions, respectively. Separate positive work reduces work by a paired median of 1.29% under the same reporting metric. Across 48 matched groups of successful square-body controllers, the median Spearman correlation between held-out field error and executed work is −0.60. Field prediction accuracy and realized control performance give different predictor rankings (Section 4, Fig. 8c, Table 7 and Appendices G.5–G.7).

Best regards,

The Authors

We have uploaded the revised manuscript. All changes in the revised manuscript are marked in red.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.