Generating highly dynamics mouvements for simulated humans and humanoids.
Auteur : Serena Ivaldi
Informations générales
| Encadrants | Pauline Maurice | Jean-Baptiste Mouret | Serena Ivaldi |
| Adresse | |||
| Téléphone | 03 54 95 85 08 | 03 54 95 86 30 | |
| pauline.maurice@inria.fr | jeanbaptiste.mouret@inria.fr | serena.ivaldi@inria.fr | |
| Bureau | C127 | C104 | C104 |
Motivation
Recent videos from Boston Dynamics show the humanoid robot Atlas performing highly dynamical motions, such as jumping, throwing boxes, running, etc., but little is known about how these movements are performed. Considering the high resemblance to human movements, our intuition is that there is an underlying form of imitation or motion retargeting, coupled with a re-optimization on the robot.
Recently, we developed real-time tele-operation for the humanoid iCub. This is possible by combining a module that retargets human demonstrations (acquired by a wearable suit from Xsens) with a multi-task QP controller that commands the robot motion. Our system enables the teleoperation (hence the production of robot movements) of the robot doing manipulation and walking, but we are far from highly dynamics behaviors. One reason is that our controller is not exploiting the dynamics in a proper way.
In this internship we address the question: “how can we make a robot generate dynamics movements starting from human demonstrations?”.
Sujet
We propose a three steps-approach that involves simulating the human and the robot.
In the first step, we gather information about how humans perform the dynamics movements, by collecting several demonstrations from the human.
In the second step, we use this prior information to optimize a simulation of the human movement, i.e., to find the optimized trajectories of the limbs/joints/end-effectors that maximize a fitness function associated to the realization of the dynamic movement. We may use stochastic optimization algorithms or reinforcement learning algorithms. We will need to formulate and implement the robot controller that executes the proper movements exploiting the dynamics of the movement. Our current QP controller being in velocity, we will need to formulate the control problem making the torques explicit.
In the third step, we may adapt the simulated human movement to the robot, following the same method of the second step but changing the fitness function and adapting the QP problem to the robot structure.
All these steps may be performed in simulation, using Gazebo and/or Dart simulators. If there is time, a demonstration on the iCub robot can also be done.
Cadre du travail
The subject builds on our current work in QP control, optimization, dynamic simulation and robot teleoperation, realized in the context of the European Project AnDy. The student will collaborate with a team of engineers, PhD and postdoc.
References
Serena Ivaldi, Lars Fritzsche, Jan Babic, Freek Stulp, Michael Damsgaard, Bernard Graimann, Giovanni Bellusci, and Francesco Nori. Anticipatory models of human movements and dynamics: the roadmap of the andy project. In Proc. International Conf. on Digital Human Models (DHM), 2017.
Penco, L.; Clement, B.; Modugno, V.; Mingo Hoffman, E.M.; Nava, G.; Pucci, D.; Tsagarakis, N.G.; Mouret, J.-B.; Ivaldi, S. (2018) Robust real-time whole-body motion retargeting from human to humanoid. Proc. IEEE/RAS International Conf. on Humanoid Robots (HUMANOIDS).
