Introduction
The need for humanoid robots
In our society, there is an increasing need for humanoid robots. This kind of robots try to reproduce several human features and integrate them in a single machine. There has been several famous examples in the past decades, as covered in details in [SK08]. Two famous examples are ASIMO, the multi-purpose research robot made by Honda, and the infant-like iCub project, developed at the Italian Institute of Technlogy (IIT) and funded by the European Commission.
Humanoid robots are a way to better understand humans. Indeed, one can use the robots to understand the principles behind the human capabilities. There are lots of areas in which that idea applies: communication, learning, physical exploration of the world, etc. In this semester project, we look at the locomotion aspect of these humanoid robots.
Humanoid robot locomotion
When we consider human locomotion, which seems very complex, we can ask ourselves the following question: ”Why didn’t the evolution give us wheels?”. Indeed, a mobile robot equipped with wheels seems much easier to control than one with legs. Therefore, legged locomotion must have some key advantages compared to wheeled locomotion.
One of these advantages is that legged locomotion allows for a better handling of rough terrains. Indeed, a wheeled robot is in trouble when faced with stairs, rubble or that kind of uneven terrain. The reason is that wheels (or caterpillars) always need a continuous support from the ground. Legs, however, allow a robot to walk even with a discrete ground support. Another advantage is that a legged robot can potentially be lighter than a wheeled one, since there is no need to have a frame to support wheels.
At this point, we have to separate between two kinds of humanoid walking. On the one hand, there is static walking, where the robot remains stable at every moment of the motion. On the other hand, there is the dynamic walking, where the robot falls if it is frozen in the middle of the motion. With this kind of gait, the stability can only be achieved using an active control of the leg motion. Therefore, dynamic walking requires to use a dynamical model of the robot whereas static walking is achievable considering only the kinematics of the robot.
In everyday life, humans walk dynamically and not statically. Dynamic walking looks much more natural than static walking. Furthermore, if the evolution favoured dynamic walking, it must be more energy-efficient than static walking. In this semester project, we focus ourselves on the implementation of a gait controller to achieve dynamic walking with a humanoid robot.
Using a muscle-reflex model to generate human-like walking gaits
Now, getting back to the introductory question for the last paragraph, we wonder whether the control of a legged walking gait is really so difficult to achieve or if there is some way to reduce this apparent complexity. It seems that the properties of the legs play a major role in the stability of the gait, as summarized in [GH10].
In the same article, Hartmut Geyer provides a controller to make a humanoid walk dynamically. It is based on a muscle-reflex model, i.e. it uses a set of contracting muscles activated by several reflex rules to produce the torques applied on the sagittal joints of the robot legs. So, it is a two-dimensional human-like gait controller.
This controller has already been implemented in simulation by previous students at Biorob, namely Steve Berger [Ber11] and Florin Dzeladini [Dze13]. Our project is about merging the reality gap between previous simulations and a real state-of-the-art robotic platform.
From simulation to reality
The main difficulty we had to cope with was to implement this model on a real robot, which brings several problems. First, the robot has a real body, with e.g. moving arms whereas the upper body of the robot in the original Geyer’s model was simply modelled by a single Head-Arms-Trunk (HAT) rigid body. Furthermore, the actual robot segments do not have the same lengths as in Geyer’s model. A scaling of the model parameters is then required. Particularly, the relative size of the real robot foot with respect to its body’s height is bigger than the one in Geyer’s model. Another major difference between the simulated version of the robot in Geyer’s model and the real one is that the real robot has actuators to apply torques whereas the ideal version assumed that the computed torque references were directly applied on the robot joints.
International context of the project
This project is part of an ongoing international collaboration. Indeed, the robot we use for this project, the CoMan, has been designed and built by the Italian Institute of Technology (IIT, Italy). It is based on the iCub, developed at IIT, too. The simulation environment used to implement the dynamic simulator of the robot has been developed at the Université catholique de Louvain (UCL, Belgium). This is the Robotran simulation environment. Finally, the Biorob Laboratory (EPFL, Switzerland) has bought the latest release of the robot and possesses a strong expertise in developing locomotion controllers for humanoid robots. Our project is shared between UCL and EPFL. This report overviews the first part of the project, carried out at EPFL during our Erasmus stay.
Results achieved
To realize this project, we first designed the gait controller on the robot model. Then, we performed some preliminary tests with this controller on the real robot. To speed up the transfer of the gait controller, developed in simulation, to the robot (and potentially vice-versa), we organized the simulator code in several layers with well-defined interfaces between them. This way, it is possible to reuse most of the controller files and transfer them without modification to the actual robot.
Report
Report of the semester project:
Results
Activation of one muscle
All the virtual Hill-type muscles receive the minimal activation, except the HFL muscle which receives a linearly increasing activation:
Perturbations on one activated muscle
The virtual HFL muscle is activated; external forces are applied to counteract the effect of this muscle:
Testing the reflexes on one foot with no contact with the ground
Two ropes at the neck of the CoMan prevent it to walk (the chain of reflexes is broken); we test the reflexes of the foot with no contact with the ground:
Walk initiation on the ground
Two ropes at the neck of the CoMan prevent it from walking, but the walk initiation is still possible; these two videos show that the CoMan decides by itself which leg (left or right) will be the first leg in swing phase:
Walk in simulation
This video shows the CoMan walking in simulation (Robotran):
Bibliography
- [SK08] Siciliano and Khatib, editors. Springer Handbook of Robotics, chapter 56. Springer, 2008.>
- [Ber11] S. Berger. Energy consumption optimization and stumbling corrective response for bipedal walking gait. Master’s thesis, Ecole Polytechnique Fédérale de Lausanne (EPFL), Biorob Laboratory, Lausanne, June 2011.
- [Dze13] F. Dzeladini. (unknown). Master’s thesis, Ecole Polytechnique Fédérale de Lausanne (EPFL), Biorob Laboratory, Lausanne, January 2013.
- [CER09] CEREM. Modeling Multibody Systems with ROBOTRAN, September 2009.
- [DMMC+12] H. Dallali, M. Mosadeghzad, G.A. Medrano-Cerda, N. Docquier, P. Kormushev, N.G. Tsagarakis, Z. Li, and D.G. Caldwell. Development of a dynamic simulator for a compliant humanoid robot based on a symbolic multibody approach. IIT, 2012.
- [GH10] H. Geyer and H. Herr. A muscle-reflex model that encodes principles of legged mechanics produces human walking dynamics and muscle activities. Neural Systems and Rehabilitation Engineering, IEEE Transactions on, 18(3):263 – 273, 2010.
- [GSB03] H. Geyer, A. Seyfarth, and R. Blickhan. Positive force feedback in bouncing gaits? Proc. R. Soc. Lond. B, 270(1529):2173–2183, October 2003.
- [GSB04] H. Geyer, A. Seyfarth, and R. Blickhan. Spring-mass running: simple approximate solution and application to gait stability. J. Theor. Biol., pages 315 – 328, 2004.
- [McM84] T.A. McMahon. Muscles, Reflexes, and Locomotion, chapter Chap. 9 – Effects of Scale, pages 234 – 247. Princeton University Press, 1984.
- [MMCS+12] M. Mosadeghzad, G.A. Medrano-Cerda, J.A. Saglia, N.G. Tsagarakis, and D.G. Caldwell. Comparison of various active impedance control approaches, modeling, implementation, passivity, stability and trade-offs. In Advanced Intelligent Mechatronics (AIM), 2012 IEEE/ASME International Conference on, pages 342 – 348, July 2012.
- [SF03] J.-C. Samin and P. Fisette. Symbolic modeling of multibody systems. Dordrecht ; Boston : Kluwer Academic Publishers, 2003.
- [SGT12] A. Schepelmann, H. Geyer, and M. Taylor. Development of a testbed for robotic neuromuscular controllers. In Proceedings of Robotics: Science and Systems, Sydney, Australia, July 2012.
- [TLSC11] N.G. Tsagarakis, Z. Li, J. Saglia, and D.G. Caldwell. The design of the lower body of the compliant humanoid robot “cCub”. In Robotics and Automation (ICRA), 2011 IEEE International Conference on, pages 2035 – 2040, May 2011.
- [TMDMC12] N.G. Tsagarakis, S. Morfey, H. Dallali, and G.A. Medrano-Cerda. COmpliant HuMANoid Platform (COMAN), 2012.