Human Motor Primitives (WP1)

Humans are an extraordinary example of autonomous systems with multiple degrees of freedom, compliant mechanics, and rich and adaptive motor skills. This work package aims at grounding the development of a novel robot architecture on the human motor control and cognitive architecture and at providing new fundamental and theoretical insights into its organization.

Leading institute: Fondazione Santa Lucia
Contact Person: Prof. Andrea d'Avella

Motion Capturing of human data

Compliant Systems (WP2)

Classical actuators found in the majority of today’s robots are stiff position controlled units for precise trajectory tracking. This work package will focus on the design, realization testing of a new generation of robotic platforms actuated by compliant electromechanical movers. The objectives of this WP are:

  • Design and realize mechatronic actuation groups with stiffness close to those found in biological systems 
  • Integrate the developed compliant actuators into a new generation of robotic systems, the humanoid iCub and the quadruped robot Cheetah 
  • Provide compliance in simulations of iCub, Cheetah 
  • Develop appropriate joint level control strategies that will allow the effective control of the joint motion and stiffness 
  • Experimentally evaluate the two robotic platforms and the control strategies developed comparatively to the non-compliant versions

Leading institute: Italian Institute of technology
Contact person: Prof. Darwin Caldwell

Leg of the cCub robot

Morphological Computataion (WP3)

In the context of robotics, the term “morphological computation” usually indicates the employment of the body as an active component in signal processing and in achieving a set of desired behaviors in general. Many studies have reported that this concept makes us control complex system such as compliant robot easily. However, a general framework for the description of morphological computations is lacking. The goals of this work-package are to build formal description of a broad variety of compliant robotic structures, to identify “extended” motor primitives in compliant systems that in addition to neural circuitry take into account the morphology of the body, and to analyze the effects of morphology on stability and on the ability to adapt to new variants of already learned tasks.

Leading institute: University of Zurich - Artificial Intelligence Laboratory

Adaptive Modules (WP4)

The objective of this work package is to design the building blocks of the complete hierarchical control architecture. The building blocks, the motor primitives, will be implemented as dynamical systems and should exhibit a set of interesting properties such as the ability to produce discrete and periodic motions,  and the possibility to be modulated in real time. Furthermore they should be suitable for learning (i.e. have open parameters that can be adjusted) and for being combined in multiple ways (e.g through superimposition and sequencing) in order to generate more complex movements. More specifically, the goals of his WP are therefore to explore the various options for designing adaptive modules using dynamical systems (both discrete and rhythmic movements),  to provide mathematical tools to analyze stability, and to provide adaptive modules as building blocks for the complete architecture and demonstrate simple locomotion and reaching skills based on a minimal number of modules.

Leading institute: EPFL
Contact person: Prof. Auke Ijspeert

Learning (WP5)

This work package researches novel learning algorithms which arise from the context of rich motor skills and the hard learning problem introduced in the experimental scenarios. The research in this work package is however of a more theoretical nature and the dissemination of the results will happen to the broader machine learning community. WP6 combines the results from this WP5 in an architecture which will be used in the actual robotic experiments. The following objectives apply:

  • Development of new learning algorithms and the redesign of existing learning algorithms for recurrent neural networks, especially exploiting reservoir computing methods.
  • Integration of different learning algorithms in complex modular control architectures.
  • Development of learning algorithms that operate in interaction with a human caretaker, by scaffolding and imitation.
  • Development of cognitive-level learning paradigms supporting agent autonomy, especially for autonomous discovery and meta-learning.

Leading institute: Graz University of Technology

 

Architectures (WP6)

The purpose of the "architectures" workpackage is to find ways to join the modules and mechanisms into comprehensive motor behavior control architectures. This can certainly not be achieved by a mere adding up of parts. A number of fundamental questions need to be resolved concerning the nature of top-down vs. bottom-up interactions between modules; the representations within modules and the communication/transformations between modules of world and body features, goals, rewards and control actions; interactions between learning mechanisms and more. We will structure the research in this WP by progressing through an evolutionary sequence of architectures, each of which lends itself to the design of a “complete Iguana” (to use a famous phrase of Rodney Brooks), rising in level of cognitive autonomy and flexibility.

Leading institute: Jacobs University Bremen
Contact person: Prof. Herbert Jaeger

Robotic Experimentation (WP7)

The main objectives of the Robotic Experimentation is to transfer the scientific innovations from other work packages into a software architecture facilitating research on rich motor skills both within the project and within the scientific community at large. For that methodological, theoretical and technological challenges from a software engineering and robotic control architecture viewpoint have to be investigated.

The realization o the finding will lead to a layered component architecture and a corresponding meta-model for the endowment of rich motor skills including high-level cognitive abilities for robotic experimentation. The developed software architecture and computational models will be proven in full body robot systems like the i-Cub and Cheetah in complex real-world tasks like crawling up a sofa and in smooth interaction with humans. We will provide feedback to other work packages about the feasibility of the developed computational models from an architecture engineering and robotic experimentation viewpoint and develop and apply evaluation methods based on human perception and cognition of motion.

Leading institute: Bielefeld University
Contact person: Prof. Jochen Steil