Adaptive Drift Compensation for Soft Sensorized Finger Using Continual Learning

TL;DR

We tackle drift in soft-sensor proprioception using an adaptive continual learning framework that couples a lightweight LSTM core with a small rehearsal memory and a regularization term to preserve the base signal while learning drift. The model is activated by drift detection via Wasserstein distance and alternates between a static base and a dynamic adaptive path to minimize forgetting. Nine drift-enriched experiments with a piezoelectric Piezola sensor on a Pneunet finger show the CL approach outperforms two benchmarks and an ablation confirms the value of the adaptive training plus rehearsal and regularization. Limitations include reliance on a replay memory; future work proposes autoencoder-based generative replay to remove memory dependence and enable drift detection.

Abstract

Strain sensors are gaining popularity in soft robotics for acquiring tactile data due to their flexibility and ease of integration. Tactile sensing plays a critical role in soft grippers, enabling them to safely interact with unstructured environments and precisely detect object properties. However, a significant challenge with these systems is their high non-linearity, time-varying behavior, and long-term signal drift. In this paper, we introduce a continual learning (CL) approach to model a soft finger equipped with piezoelectric-based strain sensors for proprioception. To tackle the aforementioned challenges, we propose an adaptive CL algorithm that integrates a Long Short-Term Memory (LSTM) network with a memory buffer for rehearsal and includes a regularization term to keep the model's decision boundary close to the base signal while adapting to time-varying drift. We conduct nine different experiments, resetting the entire setup each time to demonstrate signal drift. We also benchmark our algorithm against two other methods and conduct an ablation study to assess the impact of different components on the overall performance.

Figures (9)

• Figure 1: Soft pneumatic finger at rest and actuated state. The tip bending angle is calculated using dot product between the two vectors derived from the vision markers (yelllow, blue).
• Figure 2: The sensor signals were recorded across nine distinct experiments, with each experiment repeated for five cycles. The complete setup was restarted before each experiment to simulate the drift effects.
• Figure 3: The area under the hysteresis curve (relationship between sensor signal and the bending angle) for different experiments. The experiment with the lowest hysteresis value is selected as the base signal.
• Figure 4: The proposed architecture consists of a static part and a dynamic part with a small replay buffer for rehearsal of the previously learnt knowledge. In addition, a regularization term is added to the loss function to perform knowledge transfer.
• Figure 5: The figure shows the predicted bending angles for different experiments. On top-left, the first experiment which we consider as the base signal/experiment is plotted. In this particular case the performance of the baseline model is better than the proposed CL model.