Abstract
Soft wearable robots offer a lightweight and compliant alternative to rigid exoskeletons but suffer from structural nonlinearities like hysteresis and creep that hinder accurate joint-angle estimation. This paper presents a deep learning framework utilizing Long Short-Term Memory (LSTM) networks to compensate for these effects in a multi-degree-of-freedom (DoF) shoulder sensing suit. By processing temporal data from tendon-based sensors, the proposed Early-fusion Nonlinearity Net (ENNet) achieves high-accuracy joint tracking (RMSE < 2.52°), outperforming traditional IMU-based methods and enabling real-time teleoperation of humanoid robots.
Introduction
Soft exosuits, constructed from materials like neoprene, provide a natural and affordable interface for human-robot interaction. However, the inherent compliance of these materials introduces complex nonlinear behaviors, specifically hysteresis—where sensor readings depend on the direction of motion—and creep—where material deformation causes drift over time. These issues make traditional mathematical modeling for complex joints, such as the 2-DoF shoulder, highly inaccurate.
Methods
The researchers developed a sensing suit equipped with four tendon-based string potentiometers routed in a bio-inspired pattern across the shoulder. To map raw sensor voltages to joint angles while accounting for temporal nonlinearities, two LSTM-based architectures were evaluated:
- LNNet (Late-fusion Nonlinearity Net): Processes each sensor through an individual LSTM layer before concatenation and feeding into fully connected (FC) layers.
- ENNet (Early-fusion Nonlinearity Net): Feeds all four sensor inputs into a single, unified LSTM network simultaneously.
LSTMs were chosen for their ability to maintain a "memory" of previous states, which is essential for identifying the direction of movement and compensating for path-dependent hysteresis and material creep.
Key Results
The system was validated using a Vicon Motion Capture system for ground truth. The ENNet architecture emerged as the superior model, likely because it better captures the correlations between sensors during multi-axial shoulder movements.
- Superior Accuracy: ENNet achieved a Root Mean Square Error (RMSE) of 2.51° for azimuth and 1.33° for elevation.
- Benchmarking: This performance significantly exceeds standard Artificial Neural Networks (5.43°/3.66°) and traditional IMU-based tracking (3.00°/2.00°).
- Real-Time Performance: Deployed on an NVIDIA Jetson Nano, the model achieved an inference time of 12ms, supporting teleoperation of a NAO humanoid robot at frequencies exceeding 25Hz.
Discussion and Future Work
The study demonstrates that joint estimation in soft robotics is fundamentally a temporal problem rather than a purely spatial one. While the current prototype uses some rigid components, future iterations will move toward a fully soft architecture. Additionally, researchers aim to improve the system's "plug-and-play" capabilities, allowing it to adapt to different users and donning variations with minimal recalibration.
Summary updated