SEVENTH FRAMEWORK PROGRAMME ICT Priority Deliverable D6.2: Skill Transfer: Factors and Findings

Dissemination Level PU Public X PP Restricted to other programme participants (including the Commission Services) RE Restricted to a group specified by the consortium (including the Commission Services) CO Confidential, only for members of the consortium (including the Commission Services) This is the revised version of Deliverable 6.2 on Skill Transfer. In line with item (3) of the Y3 decision communicated by the Project Officer: (3) Drop parts which are unachievable or ineffective (workshop at autumn 2014, objectives 10 (attention-based cognitive architecture) and 11 (skill transfer)). In line with this recommendation, skill transfer was not specifically addressed further in Y4. 1 However, we briefly discuss a new experiment (in Section 3.1.2) in which skill transfer was achieved and tested as a side-effect of work conducted for D9.5, showing the domain-agnostic and embodiment-agnostic aspects of the DARWIN architecture. In this revision we address the comments from the review report (included in bold) in the following way: D6.2: The report should explain why the focus has been shifted away from skill transfer from humans to skill transfer between robots. We clarified in the introduction that, following the DoW, the topic of Deliverable 6.2 is skill transfer between tasks and robots, rather than observational learning from humans. Also, if both robots use the same DARWIN cognitive architecture, then the expectation would be that the high-level cognitive parts learned on one robot could simply be copied to the other robot, and that some learning and adaptation would only be necessary to account for different sensor equipment, affecting the afferent paths, and different kinematics, affecting the efferent paths, of the other robot. The report should explain in detail if this is the case or not, and if not, why. The document also reports about significantly varying requirements for training the PMPs for different robot kinematics. More detail should be given on how this is handled in practice. We expand on this in Section 3. As the reviewers rightly noted, high-level cognitive parts can be copied across, while afferent paths are made equal by using the same vision system on both platforms. The different kinematics do affect the efferent paths. We explain in Section 3.2 on the Neural PMP how this affects the size and connectivity of the neural networks driving the 2 platforms, but not the data gathering, training procedures, and the computational model that generates the actions. We refer throughout the document to papers …