Learning to Detect Collisions for Continuum Manipulators without a Prior Model

Shahriar Sefati | Shahin Sefati | Iulian Iordachita | Russell H. Taylor | Mehran Armand
MICCAI 2019: Medical Image Computing and Computer Assisted Intervention – MICCAI 2019
DOI: http://doi.org/10.1109/IROS.2013.6696866

The high dexterity, flexibility and compliance of the continuum manipulator (CM) make it suitable for minimally invasive surgery. CMs are normally operated in proximity of sensitive organs so it’s neccessary to have a method to detect when the collision with surrounding environment is occured.

Several methods have been proposed in the literature. However, they have their own disadvantages such as they need additional sensors, require the exact modeling of the CM or relying on geometrical assumptions and properties specific to each kind of CM. The new approach proposed by Sefati et al solely relies on data from already available sensors – Fiber Bragg Grating (FBG), which are often embedded into the catheter for shape estimation. Moreover, the model of CMs and prior assumption or knowledge regarding the geometry of CM or surrounding are unnecessary.

The problem of collision detection is defined as a supervised classification problem with two classes – collision and no collision. Sensory data captured from FBG is used as the input for the learning model and the output is the corresponding class of collision. The presented method consists of an offline dataset creation step during which, the sensory data is labeled with correct collision class. Gradient boosting classifier is trained to learn the nonlinear mapping from wavelength data collected from the FBG to the appropriate class of collision. The supervised machine learning model is optimized by tuning the hyper-parameters via a K-fold cross validation method. The performance of collision detection algorithm is evaluated on unseen test data from CM collision with different properties placed at unknown location relative to the CM. The result demonstrated successfully detection of collision on unseen data with hard and soft obstacles such as hand, soft gelatin phantom and soft sponge foam.