A deep inverse-mapping model for a flapping robotic wing

Authors: Hadar Sharvit, Raz Karl, Tsevi Beatus

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Here, we report a machine-learning solution1 for the inverse mapping of a flapping-wing system based on data from an experimental system we have developed. Our model learns the input wing motion required to generate a desired aerodynamic force outcome. We used a sequence-to-sequence model tailored for time-series data and augmented it with a novel adaptive-spectrum layer that implements representation learning in the frequency domain. To train our model, we developed a flapping wing system that simultaneously measures the wing s aerodynamic force and its 3D motion using high-speed cameras. We demonstrate the performance of our system on an additional open-source dataset of a flapping wing in a different flow regime. Results show superior performance compared with more complex state-of-the-art transformer-based models, with 11% improvement on the test datasets median loss .
Researcher Affiliation Academia Hadar Sharvit1,2,3, Raz Karl1,2,3 & Tsevi Beatus1,2,3 1 School of Computer Science and Engineering 2 The Institute of Life Sciences 3 Center for Bioengineering The Hebrew University of Jerusalem, Israel 9190401 EMAIL
Pseudocode Yes Algorithm 1 Adaptive Spectrum Layer (ASL) Require: Tensor x with shape [H, F] (or batched with shape [B, H, F]), Require: fmax, the maximal frequency to consider, and fs the sampling frequency Require: dropout rate p (0, 1) Nf [0, 1, 2, ..., H/2 1, H/2] / (H/fs) ˆx rfft(x) [:, : Nf, :] Real valued Fast Fourier Transform ˆs [|ˆx|, cos ( ˆx), sin ( ˆx)] Stacking phase and magnitude H Re LU(Fully Connected(ˆs)) Hidden representation in Fourier space w Sequential( H dropout[p=p](H) H Fully Connected(H) H H sigmoid(H) H sigmoid(H) ) Gating Mechanism ˆx Padding(ˆx w) with H Nf zeros Results in reconstruction of the same shape x irfft(ˆx)
Open Source Code Yes 1Framework, models, and data are publicly available on github.com/Hadar933/Adaptive Spectrum Layer
Open Datasets Yes 1Framework, models, and data are publicly available on github.com/Hadar933/Adaptive Spectrum Layer... The second dataset we used (Table 1, Fig. 4) has been published by Bayiz & Cheng (2021b), who measured the wing kinematics and aerodynamic forces of a plate-like wing flapping in mineral oil... Yagiz Bayiz and Bo Cheng. Flapping wing aerodynamics with prssm [dataset]. https:// datadryad.org/stash/dataset/doi:10.5061/dryad.zgmsbccbs, 2021a. URL https://datadryad.org/stash/dataset/doi:10.5061/dryad.zgmsbccbs.
Dataset Splits Yes Both datasets were randomly divided such that 75% of the events were used as a training set, 10% reserved for validation, and 15% for testing.
Hardware Specification Yes training one model took 3min on a single Nvidia RTX 3090 GPU with 24 GB RAM, and an exhaustive hyperparameter search took 50hr on the same hardware.
Software Dependencies No optimizer_name Choice of Py Torch optimizer. While the paper mentions PyTorch as the optimizer framework, it does not specify a version number for PyTorch or any other software libraries used. Thus, it does not provide specific ancillary software details with version numbers.
Experiment Setup Yes In our framework, we offer the flexibility to explore a larger set of training-related hyperparameters, as specified in Tab. 3 A.3... The parameters explored during this search included: Seq2Seq model args: embedding_size: Varied from 5 to 50 to assess its impact on feature representation. hidden_size: Varied to explore the influence of different hidden dimensions. n_layers: Investigated different numbers of RNN layers (1-3 layers). attn_heads: Explored different numbers of attention heads (1-10). batch size: Explored values ranging from a few dozen to a few thousand. feature window size: Varied (128-512) to assess sensitivity to temporal span. normalizer schemes: every possible pair of normalization schemes for the features and the targets, namely, every pair in {z-score, min-max, none} {z-score, min-max, none}... Concretely, the best-performing parameters are specified in Tab. 4... Table 4: Hyperparameters of the best models. The hyper-parameters used to train the Seq2Seq+ASL model on the validation set of each dataset