Over 500,000 Robot-Assisted Minimally-Invasive Surgeries were performed in 2014. There is a large and growing corpus of kinematic and video recordings that have potential to facilitate human training and the automation of subtasks. A key step is to segment these multi-modal trajectories into meaningful contiguous sections in the presence of significant variations in spatial and temporal motion, noise, and looping (repetitive attempts). Manual segmentation is prone to error and impractical for large datasets. We propose Transition State Clustering (TSC), which segments a set of surgical trajectories by detecting and clustering transitions between linear dynamic regimes. TSC aggregates transition states from all demonstrations into clusters using a hierarchical Dirichlet Process Gaussian Mixture Model in two phases, first over states and then temporally. After a series of merging and pruning steps, the algorithm adaptively optimizes the number of segments, and this process gives TSC additional robustness in comparison to other Gaussian Mixture Models (GMMs) algorithms. In a synthetic case study with two linear dynamical regimes, when demonstrations are corrupted with noise and temporal variations, TSC finds up to a 20% more accurate segmentation than GMM-based alternatives. On 67 recordings of surgical needle passing and suturing tasks from the JIGSAWS surgical training dataset, supplemented with manually annotated visual features, TSC finds 83% of needle passing segments and 73% of the suturing segments found by human experts

http://goldberg.berkeley.edu/pubs/ISRR2015_Krishnan-Garg_TSC.pdf