Visual sensing of continuum robot shape using self-organizing maps

Shape control of continuum robots requires a means of sensing the the curved shape of the robot. Since continuum robots are deformable, they take on shapes that are general curves in space, which are not fully defined by actuator positions. Vision-based shape-estimation provides a promising avenue for shape-sensing. While this is often facilitated by fiducial markers, sometimes fiducials are not feasible due to either the robot's application or its size. To address this, we present a robust and efficient stereo-vision-based, shape-sensing algorithm for continuum robots that does not rely on fiducials or assume orthogonal camera placement. The algorithm employs self-organizing maps to triangulate three-dimensional backbone curves. Experiments with an object with a known shape demonstrate an average accuracy of 1.53 mm on a 239 mm arc length curve.