Section: New Results
Achroban: a humanoid robot platform with dynamic balancing and semi-passive vertebral column
The AcRhoban project deals with motor primitive learning for humanoid personal robotic. The robots in which we are interested have rich but a priori quite imprecise mechanical structure. This makes difficult to get analytic models for them. Instead of that our goal is to make them learn motor behaviours such as dynamic balancing, walking on non homogeneous surfaces, and even acrobatic moves. We propose an approach mixing the use of compliant motors with hybrid position/force control and learning. This should allow flexibility and reactiveness of moves, but also a better security and energy efficiency.
At the moment, the AcRhoban project follows the following directions :
A] Design and construction. The design study has been sufficiently advanced to allow the construction of a first prototype. This prototype includes 32 servo-motors allowing a control in position, but not only. Indeed, these servo-motors allow a fine control of the parameters of their position control low-level closed loop. In particular, one can control the compliance level of the mechanism, making it close to force control. In the design of AcRhoban structure, we focused on making the torso structure particularly rich, designing a simple spine with 5 degrees of freedom. The goal is to experiment the possibilities of balancing the robot using modifications of the shape of its torso.
B] Control and Learning. AcRhoban is controled by a software environment which has been developped within the Rhoban project from several years now. This environment allows to define robot motor behaviours on the basis of :
piecewise-linear trajectories, designed by the user.
PID (closed) reaction loops, implementing the reactions of the robot relatively to its environment (e.g. inertial sensors, but also control interfaces).
Moreover, a crucial point is the interaction between these two kinds of components of moves :
trajectories can be used to define time variations of the closed loop parameters
conversely (and even recursively), reaction loops can be used to regulate some parameters of the trajectories (amplitude, shift, etc.)
This system has been upgraded and is already used to design moves of AcRhoban, including " soft " compliant moves. So, the balancing system and the walking moves include programmed fixed trajectories, but also reactions to the environment of several natures : 1) reactions in joint positions (such as the pelvis position for balancing), 2) compliance level reactions (such as the feet compliance during the walk) 3) and finally direct reactions of the mechanical structure itself (being possible thanks to compliance and softness control).
Also, it must be noticed that the moves are embedded in the electronic board of the robot (including an ARM7 microcontroler). This has has been made possible thanks to significant embedded development. The goal is to make the robot completely autonomous (i.e. without wires).
This first prototype has been showed into several public demonstrations, involving collegues, but also large public show (futuro remoto 2009). This demonstrations emphasized a crucial aspect of the project : the use of soft joints allows the physical interaction with human. On top of the intrinsic interest of the public for such interaction, this open the field of experimentation of social learning including physical contact with humanoid robots on which we will focus in the next months.