Learning by Observation for Surgical Subtasks: Multilateral Cutting of 3D Viscoelastic and 2D Orthotropic Tissue Phantoms
Main Idea: Let's design the tasks of debridement and pattern cutting as finite state machines (i.e. a set of rules with if-then steps, etc) based on human demonstrators, and then automate those tasks with dvrk.
Main Points:
-
The "finite state machine" thing is really a set of rules to follow, if we do this then do that, otherwise return to this, etc. The "learning by observation" terminology refers to using human demonstrators to update their finite state machine. The "machine" depends on precise parameters which have to be manually adjusted as needed.
-
Debridement: split into about five subtasks (not including detection). The detection part uses OpenCV and hue-saturation-value separation for detection of those parts to remove. Repeatability (i.e. success rate) of 96% in 50 trials.
-
Pattern Cutting: split into about ten subtasks. More OpenCV stuff to detect contours. Repeatability (i.e. success rate) of 70% in 20 trials.
Thoughts: I think the whole Finite-State Machine thing may be overly formal for their purposes, particularly with the debridement stuff. But maybe this is standard for such papers. (Also, Fig. 5 about the software stack is really unhelpful, not that informative.) Regarding the experimental results, I just wish there's a better way to describe it in a paper rather than regurgitating statistics derived from their tables. I would prefer more high-level conclusions from the tables. They do have a video, though, and the result looks really awesome there!! A lot of the paper is also about simply describing the rules for each step, which definitely took a lot of time and tweaking with OpenCV.
(Note: this paper was a finalist for the best medical robotics paper at ICRA 2015.)