Real-time sensing and control of multi-degree-of-freedom soft balloon actuators using data driven techniques
Embargo Date
2027-05-27
OA Version
Citation
Abstract
Soft robotic actuators can be useful in congested and limited spaces, but controlling them is difficult due to strong non-linearity, modeling complexity, and inherent stochasticity, especially underwater. Controllers based on neural networks have been proposed as one of the methods that could be applied to these robotic systems. This thesis focuses on learning-based sensing and control for an underwater Stacked Balloon Actuator (SBA) with three chambers and 25 layers. The actuator is fabricated from thermoplastic polyurethane and includes embedded copper traces that route sensor signals to the tip without noticeably affecting its motion. A synchronized setup is developed that records pump commands together with electromagnetic tip tracking and IMU/magnetometer measurements, producing a dataset on the order of 7 ×10^5 samples. Using this data, a two-stage learning open loop control pipeline is developed: a residual network predicts tip orientation (quaternion) from position differences, and another neural network predicts chamber volumes from magnetometer and orientation features. Experiments show a mean orientation error below 7.7◦, and the predicted volumes with over 99% accuracy.
Description
2026
License
Attribution 4.0 International