Sensor Contribution Analysis for Multimodal Estimation in Soft Robotic Actuators

Soft robotic systems present substantial challenges for state estimation due to their nonlinear actuation dynamics, material hysteresis, and underactuated geometries. These challenges are especially pronounced in Shape Memory Alloy (SMA) driven platforms, where mechanical motion is coupled with thermoelectric behavior that varies over time. Learning-based models such as Gated Recurrent Units (GRUs) have demonstrated strong predictive performance for soft robotic state estimation, but the role of individual sensor inputs is often unclear. This work investigates sensor contribution and redundancy within a GRU-based forward model trained to estimate the three-dimensional position of an SMA-actuated soft robotic limb. A structured ablation framework is employed to train separate models with specific sensor modalities omitted from the input data, including current, infrared temperature, and inertial measurements, enabling quantitative assessment of each modality’s impact on prediction accuracy. Results from 31 ablation configurations indicate that reduced sensor sets have the potential to outperform the full multimodal suite. Excluding the IR and PWM sensors improved the root mean squared error (RMSE) by 0.5% relative to the baseline full-sensor configuration, while reducing input dimensionality by 44%. Conversely, removing IMU sensors, particularly accelerometers, caused the greatest performance degradation, indicating their unique importance for capturing dynamic motion. Models trained with only accelerometer inputs showed 11.6% degradation compared to baseline, demonstrating that while accelerometer data is critical, it benefits from complementary sensor information. These results show that selective use of sensors can improve both estimation accuracy and computational efficiency. The proposed ablation-based interpretability framework provides a generalizable method for identifying critical sensing modalities and optimizing embedded architectures for future embedded and autonomous soft robotic systems.