UC San Diego

Continuum robots achieve high dexterity in tight spaces by bending and flexing along their entire length. Their unique modes of actuation and inherent flexibility make them ideal candidates for ``keyhole'' applications in delicate environments, such as minimally invasive surgery, where the robot must perform a useful task through a small opening without causing harm to its surroundings. The full potential of continuum robots has yet to be realized however as the advantages offered by the mechanics of these robots pose challenges for their modeling, estimation, and control.

For example, real world phenomena such as friction, material variability, and fabrication tolerances play a large role in determining the ultimate shape of a continuum robot, but are notoriously difficult to model analytically. Further, their small form factor and flexible nature make it difficult to integrate sensors that can directly measure their states. These problems are compounded when the robot comes into contact with unknown obstacles, which change the robot shape in ways that cannot be predicted a priori.

This work looks at three different algorithmic approaches to address these challenges. First, we consider how a low-parameter, numerically stable representation of the differential kinematics for a growing robot can be used together with a real time estimation technique to steer the robot through highly constrained environments. Second, we show how traditional Gaussian Process models can be adapted to produce high accuracy models for a wide range of continuum robots using data that is received sequentially over time. Finally, we develop a simplified dynamics model that facilitates the design of an observer for continuum robot control in the extreme case when the robot does not have any integrated sensing. All of our methods are validated through hardware experiments and their associated trade-offs are discussed.