Subsequent projects | Special projects

Start of funding 01.07.2008

Internal models in everyday manipulation tasks

Prof. Dr. Michael Beetz
Technische Universität München
Intelligent Autonomous Systems Group

Prof. Dr. Stefan Schaal
University of Southern California, Los Angeles
Department of Computer Science and the Neuroscience Program


A very impressive aspect of human manipulation of objects in everyday environments is that we select and parameterize our reaching and grasping movements very skillfully, which leads to smooth, predictable, and efficient motion. Furthermore, we are capable of learning these skills from little experience and adapting them automatically when needed. To do so, the brain constructs internal models of the body's interaction with the world: inverse models, which predict the motor command required to achieve a goal state, and forward models, which predict the next state based on the current motor command.

Our groups will cooperate in developing computational models of motor control involved in every day manipulation tasks, and implementing these models to control an articulated robot. The emphasis will be on studying internal models, and their role in effective manipulation in both humans and robots. The main contribution will be to study internal models in the context of everyday manipulation tasks, rather than isolated reaching tasks. In the long run, manipulation control strategies that use computational models based on empirical findings from cognitive science will facilitate human control of robot prosthetics and human-robot interaction.

As the technical facilities and the expertise of the two institutes complement each other well, we will start a cooperation to pursue our overlapping research goals. The current aim is to enable knowledge transfer between the two groups, and lay the foundation for further cooperative research projects, by an intial exchange of researchers.

Final report:
In July 2008, Dr. Freek Stulp of the IAS group made a ten-day research visit to the CLMC-Lab. From many discussion on ideas for further joint research, two concrete proposals arose. First, a proposal for a JSPS Post-doc Fellowship at the Advanced Telecommunications Research Institute International in Kyoto, Japan. This institute has a long-running and intensive cooperation with the CLMC-lab, and use similar hard- and software. Dr. Stulp visited ATR from 15.01.2009 to 15.03.2009, and worked closely with researchers at ATR and the CLMC-lab on acquiring a set of Dynamic Movement Primitives for the Humanoid robot CB-i.

The second proposal, which was also granted, was a DFG Fellowship for Dr. Stulp to conduct research at the CLMC-lab, starting August 2009. The specific aims of this research project are to learn compact models of motion primitives from recordings of human interaction, as well as robot interaction with objects, and to apply them in the context of sequences of motion primitives in everyday manipulation tasks. Compact internal models simplify motion planning with and sequencing of motion primitives, and enable the robot to react more smoothly and flexibly to disturbances. Compact models are abstractions of the high-dimensional complexity in low-level motion primitives, and provide only a few task-relevant parameters to higher levels. To achieve this goal, the following research issues will be addressed: 1) Acquiring stereotypical grasping movements on the robot, which has already partially been realized at ATR; 2) Learning compact models for these movements from human, as well as robot interaction data; 3) Sequencing and optimization of several motion primitives using only a few task-relevant parameters. The robot platform on which this research will be conducted is the CB-i, which is currently on of the most advanced humanoid robots available world-wide.