Sensorless torque estimation and control using Machine Learning in Össur's Power Knee

Abstract

Force/torque feedback is crucial in robotic prostheses to ensure safety and enable natural, intuitive movement and control. However, torque measurement relies on sensors that increase the device’s manufacturing cost and weight. This project investigates the feasibility of replacing sensor-based torque measurement in Össur’s prosthetic Power Knee (PK) with real-time torque estimation using Machine Learning (ML) models trained offline, specifically for level ground walking. The study also explores the potential to adapt this solution to a new, lighter, and more cost-effective PK device which does not utilize the original PK’s torque sensor system. Sensor-based level ground walking data at slow, self-selected, and fast speeds were collected from five trans-femoral amputees to train and test a Linear Regressor, a Convolutional Neural Network (CNN), a Gated Recurrent Unit (GRU) network, and a Hybrid architecture using Leave-One-Subject-Out Cross Validation. The Linear Regressor was further implemented online in the original PK device to evaluate battery consumption. Results showed a relative torque estimation Mean Absolute Error (MAE) of less than 2.8% of the torque range for each user and a 4.95 mW reduction in battery consumption compared to running the sensor when using the Linear Regressor. Nevertheless, all ML models exhibited relative MAEs exceeding 5% of torque range when tested on the new PK device. These findings demonstrate a robust and deployable torque estimator for the original PK device that is generalizable to different users and does not increase battery consumption compared to the sensor-based approach. Further research is required to achieve comparable performance on the new PK device.