acceptodds
Under review as a conference paper at ICLR 2027

FlapKAD-State: A State-Resolved Synthetic Dataset for Bilateral Flapping-Wing Aerodynamics

Abstract

Flapping-wing flight involves strongly coupled interactions among rapid bilateral wing motion, body dynamics, and unsteady aerodynamic responses. Physical data that reveal how these processes interact are essential for data-driven aerodynamic modeling, system identification, state estimation, reduced-order dynamics, control-oriented analysis, and scientific machine learning. We introduce FlapKAD-State, a state-resolved synthetic dataset comprising 2,000 curated rigid-wing flight episodes and 720,152 valid time steps generated using synchronized rigid-body integration and quasi-steady aerodynamic modeling. Our dataset provides comprehensive bilateral flap and wing-rotation kinematics, together with their corresponding angular velocities, local body linear velocity, body angular velocity, and independently resolved aerodynamic forces for the left and right wings. These variables provide complementary descriptions of bilateral wing kinematics, body motion, and aerodynamic response within a unified episode structure. To facilitate the use and evaluation of the dataset, we provide standardized episode-level partitions, variable-length masks, channel definitions, physical units, coordinate conventions, and data-quality documentation, together with reference benchmarks that illustrate representative learning scenarios.

Primary area:
datasets and benchmarks
Keywords:
flapping-wing aerodynamics, physical dynamics datasets, aerodynamic modeling, coupled dynamical systems

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_zQVhOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a simulated flapping-wing flight dataset comprising 2,000 trajectories with synchronized recordings of bilateral wing motion, body motion, and the vertical aerodynamic force on each wing. It defines three task families: bidirectional wing-motion–force prediction, aerodynamic force prediction with additional body-state information, and body-motion prediction. These tasks examine whether observing additional physical quantities improves prediction. The results show that some information is useful, but its benefits depend strongly on the model and prediction horizon.

Critical strengths

This is a very interesting topic, and datasets and benchmark tasks for flapping-wing flight appear relatively scarce.

The three task families are reasonably designed and fairly representative, with corresponding baselines.

Critical weaknesses

The dataset appears limited in diversity. Experiments with different wing shapes are not reported, and the evidence does not sufficiently establish whether the dataset's diversity and size are adequate.

The value of the simulated data is unclear to me, and the simplified simulation model raises questions about its physical realism.

The practical value of the tasks is not clearly explained. I see limited engineering or scientific value, ( although I am not very familiar with this field.)

Questions for authors
  • Could you explain why the state-only GRU performs substantially worse than persistence for one-step body-motion prediction? In Table 5, their MSE values are 0.605 and 0.025, respectively, making the GRU's MSE approximately dozens of times as high.

  • does the simulation–experiment comparison match physical durations as well as the number of time steps?

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

This is a dataset and benchmark paper. The paper presents FlapKAD-State, a synthetic dataset of 2,000 episodes (720,152 steps at 50 Hz). It is generated by a Unity rigid-body simulator of a flapping vehicle with rigid wings, with a quasi-steady aerodynamic formula evaluated at one sampling point per wing. Each step gives a 14-dimensional state (left and right flap and twist angles and their rates, body linear velocity in the local frame, body angular velocity in the world frame) and the vertical aerodynamic force of each wing. Three benchmark families share one forecasting setup (256 steps of history, predicting 1, 16 or 32 steps ahead): forward and inverse mapping between wing motion and force, how much wing rates and body state help predict force, and how much force history helps predict body motion.

Critical strengths

Benchmark: careful data handling and evaluation protocol. The paper uses splits by episode, validity masks, normalization from training data only, windows that never cross episode boundaries, metrics balanced across episodes, and seed-paired comparisons for some models. This is good practice for sequence benchmarks and is the main positive point of the submission.

Critical weaknesses

Dataset:

  1. The physics step and the logging step are both 0.02 s. With wingbeats of about 4 to 5 Hz (my estimate from the paper's statement that the 320 ms and 640 ms horizons span 1.3 to 1.5 and 2.7 to 3.1 wingbeats), that is roughly 11 samples per wingbeat cycle, and at the peak wing rates of about 1,000 degrees per second the wing moves about 20 degrees per step.

  2. Only the vertical wing forces are released. There is no thrust, drag or sideways force, no moments, no attitude and no control input, even though body angular velocity at times reaches 7 to 8 rad/s with almost symmetric wing motion.

  3. Physical consistency has not been sufficiently analyzed.

Benchmark:

  1. The key conclusions rest on reference models that look like they were not trained properly. In B3, predicting one step ahead, the GRU has an error of 0.605, against 0.030 for simply repeating the last value and 0.018 for ridge regression. It is 20 times worse than copying the last observation.
Questions for authors

Is the wing motion prescribed or solved dynamically? Could you provide more details about your engine and settings about motion and physical attributes?

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

I'm not sure if this is a machine learning paper. It basically introduces , and evaluates on Ridge, GRU, TSMixer, Grouped-State MLP, TimeXer, PySINDy, and Neural ODE without even citing them (in experiments) or evaluating on all of them all the time.

Critical strengths

Such a dataset was lacking and I'm sure the specific community that this paper serves will appreciate. Though, the wings are like articulated objects and as far as vision/robotics methods are concerned there are comparable datasets- which this paper misses out on.

Critical weaknesses
  • I see the state representation and the benchmark families. But the main paper has to provide the full details of the simulator here. What is the exact mechanism/dynamics? What are the intricacies that separate this dataset from any other articulated benchmarks?
  • Why does the paper state 'synchronization'? aren't we collecting this data in a simulated environment? isn't everything synthetic anyway? what do the authors synchronize?
  • Rather than establishing a suitable baseline, the paper just evaluates Ridge, GRU, TSMixer, Grouped-State MLP, TimeXer, PySINDy, and Neural ODE. Also PySINDy doesn't appear in the tables of the main paper.
  • TSMixer seems to be the strong method. I expect it to appear in all experiments. Although, since the paper doesn't propose a specific baseline, that's less of an issue. Still, important to establish a leaderboard.
  • Fig. 5 is reference in Sec. 6 - it is hard to understand why it appears in the location that it is now.
  • Can the paper test shorter histories as well?
  • This paper really needs error bars. They are missing allover.

I'm not very familiar with this field. So I'll ask: In general, the experiments seem to demonstrate numerical consistency rather than aerodynamic accuracy. Is it possible to have a physical check?

Questions for authors

See weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.