We develop a control-oriented learning framework that combines a nonlinear static model with learned residual dynamics for real-time 3D control of soft pneumatic actuators. Experiments achieve ~1 mm RMSE at 10 mm/s and <10 mm RMSE at 100 mm/s, while enabling stable user-guided motion exceeding 25 m/s² peak acceleration and real-time obstacle avoidance. The learned model also predicts closed-loop stability, successfully distinguishing stable, marginally stable, and unstable behavior prior to deployment.

We extend a planar discrete elastic rod model for real-time dynamic task-space control of soft pneumatic actuators. The sparse formulation efficiently handles holonomic and nonholonomic constraints, enabling a 10-rod model and observer to run at 1 kHz while achieving ~2 mm RMSE on precision trajectories and stable tracking up to 2 Hz and 37 cm/s. We demonstrate model generalizability across three different actuator geometries, real-time tracking of user-generated references, and observer-based control that reduces tip-tracking error by 45% on average without direct tip sensing.

We develop a constrained Lagrangian model to capture the highly dynamic, non-holonomic based locomotion of a pneumatically actuated wheeled soft robot. The robot achieves speeds up to 22 cm/s (1.38 body lengths/s) and demonstrates forward, reverse, and circular locomotion. We validate the modeling framework across three different robot designs, accurately reproducing their experimentally observed configuration and velocity trends.

We develop a data-driven modeling framework that combines a nonlinear feedforward map with EDMDc residual dynamics for control of soft pneumatic actuators. Experiments achieve 2 mm RMSE on shape-rich trajectories and <10 mm RMSE at tip speeds exceeding 100 mm/s. The learned model also enables closed-loop stability prediction without requiring an explicit first-principles dynamic model.

We develop an impedance-based contact sensing method that enables a concentric tube robot to collect tissue surface points for robot-to-anatomy registration during endoscopic procedures. Experiments in an anthropomorphic prostate phantom achieve a mean surface registration error of 2.11 ± 1.01 mm, demonstrating contact sensing as an additional source of information for improving situational awareness in robot-assisted surgery.