Embedded Files
Accurate control of soft robots is critical for positioning, locomotion, and manipulation applications. These soft systems, however, are currently limited by the lack of soft multi-axis sensor designs as well as a lack of models for interpreting complex soft sensor signals. In this work, we demonstrate a unique scheme for embedding soft carbon black strain sensors into a soft multi-axis pneumatic bellows actuator. In order to utilize the soft strain sensors for estimating orientation, the strain signals from each of the 12 sensors are compared against signals from a reference inertial measurement unit. Then, a time-based recurrent neural network is used to map the hysteretic soft sensor data to determine the yaw, pitch, and roll motions of the actuator. Results show that high accuracy estimates are achieved with average bias errors between 0 to 1.5 degrees and error standard deviations between 3 to 6 degrees. This novel application of soft multi-axis embedded sensors and neural network models demonstrates how fully soft actuators and sensors can be used to accurately determine robot orientation.
The step response of a single bellows is illustrated as it is inflated and deflated. The three-axis orientation from the IMU and four strain sensor outputs are shown to illustrate the complex behavior.
The model architecture of the LSTM neural network is shown. Sensor data (X) with a dimension (d) of 114 points, corresponding to 19 samples looking back 5 steps, is input into the model. Each input passes the cell state (C) and hidden state (h) to the successive input. The intermediate LSTM output, d = (50,1), is passed into a dense function to reshape the array and match the three-axis orientation of the robot.
N. Kohls, K. Gibson, A. Singla, R. Balak, A. Jargalsaikhan, B. Bartolek, V. Gattani, Y. Mazumdar, “Carbon Black Sensor and Neural Network Model for Sensing Angle in Soft Pneumatic Actuators,” AACC Modeling, Estimation, and Control Conference, vol. 55, no. 37, pp. 199–204, 2022 [https://doi.org/10.1016/j.ifacol.2022.11.184]
Presented at AACC Modeling, Estimation, and Control Conference 2022, Jersey City, New Jersey, October 3-5, 2022
Page updated
Google Sites
Report abuse