As part of the CICADA project, a number of researchers in the Schools of Electrical and Electronic Engineering and Mathematics are working on learning and control approaches to make the iCub bipedal humanoid robot walk as well as looking at fundamental research questions. This is done in collaboration with the team at IIT (Italian Institute of Technology), Genoa.
Inherently, flexible, dynamic locomotion is a hybrid control problem with different gait phases (single support, double support, toe push off, ...) being represented by different dynamic models and controllers.
The iCub is the humanoid robot which was developed by Profs Darwin Caldwell and Giulio Sandini and their colleagues on the EU RobotCub project between 2004 and 2010. The iCub was originally developed as a platform for studying cognition, indeed it is the recommended EU testbed for such studies. A range of behaviours was demonstrated at the final review (crawling, attention seeking, drumming, ... although learning to walk was never included in the original project description. Full descriptions of the iCub are at: RobotCub Project, Wiki: RobotCub and Friends, iCub on YouTube and iCub (humanoid robotics) at IIT. Currently, the iCub is being further developed to include compliance.
Our work on the iCub involves developing hybrid control strategy for flexible, dynamic walking.
As part of the work at Manchester, we're building an accurate dynamic model of the iCub for walking simulation and control design. This is being done using Robotran. Robotran has been extended to include actuator models (electrical motor drives) and also compliant joints and a full 12 dof model of the lower body iCub has been implemented. This is currently being validated and used for dynamic analysis control design.
In addition we have developed a Maple simulation of iCub to improve the walking visualization.The humanoid model in Maple also includes the actuator dynamics, compliance and ground reaction forces in a simpler graphical interface. A case study is done with Maple called "Learning to walk faster with high performance modelling". Further details about this work is available here.
Initially, classical control techniques are being implemented and their robustness to disturbances analyzed. This has involved comparing multivariable control with classical PID link approaches to assess tracking, disturbance rejection and stability margins. It has also covered double support controller design using constraint subspaces. In addition, torque analysis is being performed to understand the range of motion achieveable using both the original and the compliant iCubs.
Some simple, initial results are available on YouTube:
While a lot of work has been done on the modelling, control analysis and design, we are still working on demonstrating the iCub walking. However, the recent work on adding compliance into the iCub has shifted the focus to demonstrating walking for the new compliant iCub.
2010.60: H. Dallali, M. Brown and B. Vanderborght Using the Torso to Compensate for Non-Minimum Phase Behaviour in ZMP Bipedal Walking. German workshop on robotics, Braunschweig, Germany ,9-10 June, 2009.
2010.69: H. Dallali, G.A. Medrano-Cerda and M. Brown A Comparison of Multivariable & Decentralized Control Strategies for Robust Humanoid Walking. UK Automatic Control Conference (UKACC), Coventry, 7-10 September, 2010. Presentation
G. A. Medrano-Cerda, H. Dallali, M. Brown, N. G. Tsagarakis, D. G. Caldwell Modelling and Simulation of the Locomotion of Humanoid Robots. UK Automatic Control Conference (UKACC), Coventry, 7-10 September, 2010.
O. Tutsoy, M. Brown Convergence Analysis of Reinforcement Learning Approaches to Humanoid Locomotion UK Automatic Control Conference (UKACC), Coventry, 7-10 September, 2010.
Speaker : Professor Bruno Siciliano
In addition, we have a reading group looking at "falling robots" (including Prof Paul Glendinning) and we're also developing links with psychologists (Dr Warren Mansell) and reseachers building "artificial brains" (Prof Steve Furber).