Biological inspiration within the STIFF-FLOP robot aims at learning from the octopus strategies to control the highly redundant arm. Motions such as reaching for food can be used to test the representation capability of the developed model, which could then be re-used to encode other types of skills within the surgical environment.
The first step was to create a procedure that is able to transfer skills between the two different embodiments and is capable of faithfully reproducing observed octopus movement. For this purpose, a set of data from different reaching movements performed by different animals was collected and analyzed by HUJI and exploited by IIT to create a statistical model of the movements. This motion was then transferred to the STIFF-FLOP robot, by exploiting the modular structure of the arm and a self-refinement procedure that improves the movement, by taking into account some predefined reward functions.
This result shows that the statistical model, could be used to reproduce a very similar movement on the STIFF-FLOP robot after a short self-refinement learning process. The next step will be that of exploiting the developed encoding strategy to perform different motions that are useful for the surgical application. The key aspect of the proposed approach is the possibility of reusing parts of the skill that have been learned for a different category of tasks, by exploiting the correlations and synergies extracted from the movement.
Qualitative comparison of the octopus and STIFF-FLOP robot movements, before (a) and after (b) self-refinement.