Despite much advancement in surgical robots, an increasing number of challenges are still open. One challenge is how to go from industrial-like robot arms that move rigid surgical tools to flexible structures that could safely move between soft tissues and freely move without the present constraint of a fulcrum point defined by the inserted trocar.

Here, the authors assume the existence of such a soft robot, a snake-like system with a number of equal segments, and address the problem of how to control its movements to follow the surgeon’s commands.

The bulk of the paper concerns the design of a learning method able to transform paths defined for a seven-degrees-of-freedom (dof) arm to the soft robot having 15 dof. The robot arm is considered as a teleoperated interface for the surgeon, which demonstrates target points to reach, and via points for the soft robot body to go through. Since the two robots have different structures, high-level imitation learning is proposed. The implementation uses a Gaussian mixture model to learn a context-reward mapping for a given set of primitives, starts a policy as a context-action mapping, and optimizes the policy in iteration loops. The simulations show how the soft robot learns to exploit its redundancy to control the whole body.

In principle, the method applies to any pair of robots, so it can be a way to transfer skills from one robot to another despite their different kinematics. This is an important step in robotics. In surgery, it is quite an impractical case, since the surgeon more easily uses a virtual interface.

The videos posted on www.vimeo.com by the first author will help the reader appreciate the project and the results.