Growing robot sensing and control
Growing robots (or vine robots) that achieve locomotion by extending from their tip, are inherently compliant and can safely navigate through constrained environments that prove challenging for traditional robots. However, the same compliance and tip-extension mechanism that enables this ability, also leads directly to challenges in their shape estimation and control.
We have developed a low-cost, wireless, permanent magnet-based method for localizing the tip of these robots. A permanent magnet is placed at the robot tip, and an array of magneto-inductive sensors is used to measure the change in magnetic field as the robot moves through its workspace. We develop an approach to localization that combines analytical and machine learning techniques and show that it outperforms existing methods.
We have developed an approach for enabling tactile perception for these growing robots. Our approach uses a set of flexible sensors that measure the curvature of the robot shape at multiple locations. We designed a pouch to enable seamless integration of the sensors with the material of the growing robot, and we developed an algorithm for determining the location of point contacts along the robot body. We evaluate the ability of the proposed approach to estimate the contact locations when the robot is grown straight, as well as when it is actively bending. We also characterize the minimum distance required to discriminate between two separate contact points along the robot body.
We have also developed an approach for controlling steerable versions of these vine robots that is based on leveraging real-time position and orientation measurements. We use the pose information to update a local model of the velocity kinematics of the robot via corrective rotations and magnitude adjustments. We evaluate the performance of this control approach in point-reaching tasks in both unconstrained and constrained environments.
Many robotics problems involve phenomena that are difficult to model from first principles. We have developed an approach that is nonparametric, models uncertainty, and requires minimal hand-tuning. Our proposed algorithm offers a combination of benefits unavailable from other online learning algorithms, including the simultaneous online optimization of local model parameters, a clear and interpretable data partitioning strategy, and principled fusion of multiple model predictions without the need for joint training. In addition to comparing the performance of our algorithm to other online and offline modeling approaches in simulation, we also demonstrate how it can be used for online learning tasks that involve physical hardware and dynamic environments. We demonstrate online learning of the kinematics of a concentric tube robot in free space and in the presence of dynamic environmental constraints.
