Haptic Rendering of Wearable Interfaces using Machine Learning

TY - CONF

T1 - Haptic Rendering of Wearable Interfaces using Machine Learning

AU - Chinello, Francesco

AU - Olsen, Martin

AU - Pacchierotti, Claudio

PY - 2020

Y1 - 2020

N2 - Wearable haptics is becoming increasingly popular in the field, also thanks to the high availability of fast-prototyping technologies as well as their promising applicability in ubiquitous immersive experiences. Of course, applying precise and timely haptic sensations is a fundamental feature of any haptic system, including wearables. This work presents an easy-to-implement approach for calibrating a position-controlled wearable haptic device using machine learning. It uses an external force sensor to create a mapping between force/torque applied by the end-effector and its pose. Using this experimental setup, we collected 2197 position-wrench pairs. With these data, we develop a machine learning rendering algorithm using standard feed-forward NNs. The objective is to design an NN acting like the inverse of the device, receiving an input force and generating a corresponding output pose of the end-effector. This constitutes a classic regression task within the field of supervised learning. We train 31 neural network (NN) able to generalize this wrench-position mapping to a broader set of commands: 11 NNs with a single hidden layer having a number of neurons in {2, 3, 4, 5, 10, 15, 20, 25, 30, 35, 40}, and 20 NNs with two hidden layers where the number of neurons in the first layer and the second layer is represented by a pair in {20, 25, 30, 35, 40}×{10, 20, 30, 40}. We test the proposed rendering approach with a 3-DoF wearable device for the fingertip, but a similar technique can be employed for controlling a large set of haptic devices, including standard single-point grounded interfaces. We analyze the performance of the different NNs in rendering forces, analyzing the relationship between their capacity and the rendering precision. Results show an error close to 0.01 N and 0.008 Nm. Increasing the number of neurons seems to have a small positive effect on the rendering performance.

AB - Wearable haptics is becoming increasingly popular in the field, also thanks to the high availability of fast-prototyping technologies as well as their promising applicability in ubiquitous immersive experiences. Of course, applying precise and timely haptic sensations is a fundamental feature of any haptic system, including wearables. This work presents an easy-to-implement approach for calibrating a position-controlled wearable haptic device using machine learning. It uses an external force sensor to create a mapping between force/torque applied by the end-effector and its pose. Using this experimental setup, we collected 2197 position-wrench pairs. With these data, we develop a machine learning rendering algorithm using standard feed-forward NNs. The objective is to design an NN acting like the inverse of the device, receiving an input force and generating a corresponding output pose of the end-effector. This constitutes a classic regression task within the field of supervised learning. We train 31 neural network (NN) able to generalize this wrench-position mapping to a broader set of commands: 11 NNs with a single hidden layer having a number of neurons in {2, 3, 4, 5, 10, 15, 20, 25, 30, 35, 40}, and 20 NNs with two hidden layers where the number of neurons in the first layer and the second layer is represented by a pair in {20, 25, 30, 35, 40}×{10, 20, 30, 40}. We test the proposed rendering approach with a 3-DoF wearable device for the fingertip, but a similar technique can be employed for controlling a large set of haptic devices, including standard single-point grounded interfaces. We analyze the performance of the different NNs in rendering forces, analyzing the relationship between their capacity and the rendering precision. Results show an error close to 0.01 N and 0.008 Nm. Increasing the number of neurons seems to have a small positive effect on the rendering performance.

KW - Wearable Haptics

KW - Machine Learning

M3 - Poster

T2 - EuroHaptics

Y2 - 6 September 2020 through 9 September 2020

ER -