Addressing data imbalance in Sim2Real: ImbalSim2Real scheme and its application in finger joint stiffness self-sensing for soft robot-assisted rehabilitation

Zhongchao Zhou, Yuxi Lu, Pablo E. Tortós-Vinocour, Ruian Qin, Shota Kokubu, Fuko Matsunaga, Qiaolian Xie, Wenwei YuFrontiers in Bioengineering and Biotechnology 2024 · DOI

Abstract

The simulation-to-reality (sim2real) problem is a common issue when deploying simulation-trained models to real-world scenarios, especially given the extremely high imbalance between simulation and real-world data (scarce real-world data). Although the cycle-consistent generative adversarial network (CycleGAN) has demonstrated promise in addressing some sim2real issues, it encounters limitations in situations of data imbalance due to the lower capacity of the discriminator and the indeterminacy of learned sim2real mapping. To overcome such problems, we proposed the imbalanced Sim2Real scheme (ImbalSim2Real). Differing from CycleGAN, the ImbalSim2Real scheme segments the dataset into paired and unpaired data for two-fold training. The unpaired data incorporated discriminator-enhanced samples to further squash the solution space of the discriminator, for enhancing the discriminator's ability. For paired data, a term targeted regression loss was integrated to ensure specific and quantitative mapping and further minimize the solution space of the generator. The ImbalSim2Real scheme was validated through numerical experiments, demonstrating its superiority over conventional sim2real methods. In addition, as an application of the proposed ImbalSim2Real scheme, we designed a finger joint stiffness self-sensing framework, where the validation loss for estimating real-world finger joint stiffness was reduced by roughly 41% compared to the supervised learning method that was trained with scarce real-world data and by 56% relative to the CycleGAN trained with the imbalanced dataset. Our proposed scheme and framework have potential applicability to bio-signal estimation when facing an imbalanced sim2real problem.

Introduction

In robotics and bio-signal processing, models are frequently trained in simulation environments because real-world data collection is often expensive, risky, or logistically difficult. This reliance creates the Sim2Real problem: models that perform optimally in a simulator often fail when deployed in the complex, noisy real world.

While unsupervised domain adaptation tools like CycleGAN (Cycle-consistent Generative Adversarial Networks) help bridge this gap, they struggle significantly with data imbalance. In many practical scenarios, researchers have access to millions of simulation points but only a handful of real-world samples. For regression-type problems—where the goal is to predict continuous values like stiffness rather than discrete categories—this imbalance leads to weak discriminators and indeterminate mappings, where the generator learns physically nonsensical transformations.

The ImbalSim2Real Scheme

The proposed ImbalSim2Real scheme departs from standard unsupervised domain adaptation by splitting the training process into two distinct "folds" to handle paired and unpaired data differently.

Two-Fold Training and Space Squashing

The dataset is segmented into paired data (where simulation and real-world conditions match exactly) and unpaired data. To strengthen the discriminator despite scarce real data, the authors introduce Discriminator-Enhanced Samples (DES). By providing a wider variety of "fake" samples, they "squash" the discriminator's solution space, forcing it to become more discerning. This, in turn, pushes the generator to produce more realistic and accurate outputs.

Targeted Regression Loss (TRL)

For the paired data, a specific Targeted Regression Loss (TRL) term is integrated. This loss directly penalizes the difference between the transferred simulation data and the ground-truth real data. This acts as a physical "anchor," ensuring the mapping is quantitative and specific rather than just visually similar.

Comparison of specific sim2real mapping issues (B) versus the corrected mapping using Targeted Regression Loss (D).

As illustrated in the figure above, standard mapping (B) often results in incorrect alignments between simulation (S) and real-world (R) domains due to the lack of constraints. The ImbalSim2Real scheme (D) utilizes TRL to ensure a "Correct" mapping, effectively anchoring the generator's output to known real-world values.

Experimental Validation

The researchers validated their approach using two distinct setups:

  1. Numerical Benchmarks: Three groups of synthetic mathematical functions representing different levels of mapping complexity (e.g., complex-to-complex, simple-to-complex).
  2. Soft Robotic Finger: A real-world application involving a soft pneumatic actuator used for rehabilitation. Simulation data was generated using Finite Element Method (FEM) modeling, while real-world data was collected from a physical prototype using a torsion spring to simulate joint stiffness.

Key Results

  • Numerical Superiority: In synthetic tests, ImbalSim2Real reduced the Validation Mean Squared Error (MSE) by 3.8x to 29.5x compared to standard supervised learning.
  • Stiffness Sensing: In the soft robot application, the framework reduced estimation error by 41% compared to supervised learning (trained on scarce data) and by 56% compared to a standard CycleGAN.
  • Success Rate: The model achieved an 88% success rate (predictions within 10% of the true value), significantly outperforming CycleGAN (41%) and supervised learning (63%).

Conclusion and Future Work

The ImbalSim2Real scheme demonstrates that even a small amount of paired data, when treated with a dedicated regression loss, can significantly stabilize GAN training for regression tasks. By making the discriminator's task more difficult through DES and providing physical anchors via TRL, the model overcomes the indeterminacy common in imbalanced datasets.

Future research will focus on establishing the minimum ratio of paired data required for optimal convergence and improving the model's sensitivity to outliers at the extreme ends of the data range. The scheme holds significant promise for high-dimensional bio-signal estimation and complex robotic sensing tasks.

Summary updated

Abstract

The simulation-to-reality (sim2real) problem is a common issue when deploying simulation-trained models to real-world scenarios, especially given the extremely high imbalance between simulation and real-world data (scarce real-world data). Although the cycle-consistent generative adversarial network (CycleGAN) has demonstrated promise in addressing some sim2real issues, it encounters limitations in situations of data imbalance due to the lower capacity of the discriminator and the indeterminacy of learned sim2real mapping. To overcome such problems, we proposed the imbalanced Sim2Real scheme (ImbalSim2Real). Differing from CycleGAN, the ImbalSim2Real scheme segments the dataset into paired and unpaired data for two-fold training. The unpaired data incorporated discriminator-enhanced samples to further squash the solution space of the discriminator, for enhancing the discriminator's ability. For paired data, a term targeted regression loss was integrated to ensure specific and quantitative mapping and further minimize the solution space of the generator. The ImbalSim2Real scheme was validated through numerical experiments, demonstrating its superiority over conventional sim2real methods. In addition, as an application of the proposed ImbalSim2Real scheme, we designed a finger joint stiffness self-sensing framework, where the validation loss for estimating real-world finger joint stiffness was reduced by roughly 41% compared to the supervised learning method that was trained with scarce real-world data and by 56% relative to the CycleGAN trained with the imbalanced dataset. Our proposed scheme and framework have potential applicability to bio-signal estimation when facing an imbalanced sim2real problem.
BibTeX
@inproceedings{zhou2024addressing,
  author = {Zhongchao Zhou and Yuxi Lu and Pablo E. Tortós-Vinocour and Ruian Qin and Shota Kokubu and Fuko Matsunaga and Qiaolian Xie and Wenwei Yu},
  title = {{Addressing data imbalance in Sim2Real: ImbalSim2Real scheme and its application in finger joint stiffness self-sensing for soft robot-assisted rehabilitation}},
  booktitle = {Frontiers in Bioengineering and Biotechnology},
  year = {2024},
  doi = {10.3389/fbioe.2024.1334643},
}
Figure