direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Page Content

Compensating the Pneumatic Actuation Noise by Using Strain and Acoustic Signals

Master Thesis

Veronika Pavlova

We are attempting to combine the liquid metal strain sensors (top) and the active acoustic sensor (bottom) to achieve improved robustness and increased measuring accuracy.


To estimate the robot state and environment, we can use different sensors. For example, we can measure the sound inside of soft pneumatic actuators and use small changes in the sound signal to sense a wide range of actuator properties (Active Acoustic Sensor). Alternatively, strain sensors attached to the robot finger  measure the deformation of the soft hull (Strain Sensor). While each sensor has shown promising results in different tasks, some disturbances lead to data corruption or loss. To achieve a robust sensor prediction, we combine multiple sensors. Additionally, this fusion might also give us more information about the pneumatic actuator state and environmental interaction and may thereby lead to a more precise contact location detection.

Description of Work

We developed two primary sensors in the lab, which we are using for different robotics tasks.
First, we have the Active Acoustic Sensor, which has a speaker and microphone attached inside at the tip and bottom chamber. The speaker generates a sound, and if the finger deforms or makes contact, the recorded sound of the microphone varies. These changes can give us precise predictions on the finger state, e.g., the applied force, deformation, or even temperature. However, one disadvantage are external sound sources, like the inflation/deflation sound, which corrupt the recordings.

Secondly, we have four strain sensors at the top and sides of the soft hull. They measure deformations of the finger via changes in the electrical resistance. The sensors are filled with liquid metal which disconnects sometimes and lead to data loss.

The goal is to combine the acoustic and strain sensors and achieve the same (or even better) contact location accuracy during all times and variable tasks. Maybe even go further and use the higher amount of state information from both sensors to, e.g., classify different objects by attributes like the density or shape. 

Zusatzinformationen / Extras

Quick Access:

Schnellnavigation zur Seite über Nummerneingabe