Seminar

Abstract:

Continuum robots with soft compliance feature are expected to be used in minimally invasive surgery. The continuum manipulators are usually small size, less constrained and accessible to the confined surgical cavity. However, accurate motion planning and stable control of those continuum robots, especially for tasks in long-time horizon planning, become extremely challenging as their accurate kinematics models are usually complicated and analytically inaccessible. To deal with the dilemma, imitation learning has been introduced for building continuum robot’s kinematics, the mapping between the actuation space, task space, and configuration space. Imitation learning is a data-driven approach based on demonstration data. Users provide demonstrations to show a specific task, then the consistency of the many demonstrations would be encoded into a statistical model and the meaningful information would be extracted out. Imitation learning directly aims at the search space for policy learning. Imitation learning enables us to understand the mechanisms between motor control and task performance. Based on those, imitation learning is a potential great method to encode inverse kinematics for continuum manipulators.

There is still limitation for imitation learning in decoding a sequential task. Motion planning for a long-time-horizon task would be much more complicated and additional technical supplements are required. Support Vector Machine (SVM) is introduced to achieving high-dimensional data classification, doing segmentation to extract the motion primitives. To model the tasks lasting for a period of time, it is reasonable to segment the task data into various motion primitives and then analyze them. Moreover, another model for sequencing the primitives is required based on probability theories. Hidden Markov Model (HMM) is that kind of model for achieving the transfer among a number of sates according to a determined topology.  Finally, a hybrid SVM/HMM is designed to build an accurate task planning model for the continuum manipulators. In conclusion, my research objectives are mainly two points: greatly improve the applicability of imitation learning; achieve modeling long-time-horizon tasks for continuum manipulators.