Full- day Workshop on Online and offline optimization for humanoid robots. at IEEE IROS 2013 in Tokyo

Transcription

1 Full- day Workshop on Online and offline optimization for humanoid robots at IEEE IROS 2013 in Tokyo Organizers: Eiichi Yoshida, Katja Mombaur, Tom Erez, Yuval Tassa Nov 7, 2013 TALK ABSTRACTS Tamim Asfour (Karlsruhe Institute of Technology (KIT), Institute for Anthropomatics, Germany) Learning and Executing Goal- Directed Motion Primitive for Humanoid Robots In this talk, we present advances on learning goal- directed motion primitives from human observation based on a reference model of the human body the Mater Motor Map (MMM) that decouples motion representation from the execution on a real robot and thus builds a unifying interface for transferring motor knowledge between different embodiments. Furthermore, we present a novel approach for automatic segmentation of human motion based on spatial object- hand relations. The resulting action sequences are executed based on previously learned library, which incorporate actions, associated objects, preconditions and effects. We demonstrate results in the context of kitchen scenario. Adrien Escande (CNRS- AIST JRL Tsukuba, Japan) Hierarchical least squares: principle and extension The notion of hierarchy of tasks, as a set of linear constraints to be solved according to some priorities, is a common concept in robotics, mostly used for inverse kinematics and control. For long, the generic resolution schemes were only able to handle equality constraints with possibly an ad hoc management of inequality constraints at the top or lowest priority level. In 2009, the first general scheme for both equalities and inequalities was introduced, using a sequence of least square programs. We then proposed an efficient numerical implementation of it, based on a tailored matrix decomposition, the Hierarchical Complete Orthogonal Decomposition (HCOD), and an adaptation of the active set method for least squares. The obtained algorithm proved to be much faster than the classical SVD approach in the equality- only case, and fast enough for online use in the inequality case. In this talk, we will review the approach and discuss some interpretations and uses of the HCOD. We will then introduce the latest developments obtained thanks to the algorithm. Jean- Paul Laumond (LAAS- CNRS, Toulouse, France) The geometry of human locomotion The talk will overview a current research aiming at better understanding the role of the head in the control of bipedal locomotion. In a first part we see how human vision induces trajectory asymmetry in SO(2). In a

2 second part we show how a top- down approach of locomotion control relates to the notion of textured terrains and how it opens questions about cognitive locomotion architectures. Dinesh Manocha (University of North Carolina, USA) Real- Time Motion Planning for High- DOF Robots and Sensor Data Algorithmic motion planning has been actively studied in robotics and related areas for more than three decades. There is a rich collection of techniques that have been successfully used for CAD/CAM, bioinformatics, computer gaming and other applications. However, current techniques have two major limitations in terms of applications to physical robots. They are mostly designed for static environments or assume an exact geometric representation of all the obstacles in the scene. In this talk, we give a brief overview of our recent work on motion planning for high- DOF robots. This includes two new planning algorithms based on optimization- formulation as well as sample- based planning. The optimization formulation can take into account collision- free and dynamics constraints, and handles dynamic obstacles using a replanning framework. In terms of sample- based planning, we present a novel RRT- based algorithm based on Poisson sampling. Both these planners can be parallelized on commodity, many- core GPUs and offer considerable speed over prior techniques. We highlight their performance on human- like models with 26- DOF. We also present novel techniques to perform collision and proximity queries with point- cloud sensor data, collected using depth or other sensors. This includes novel techniques for robust collision detection using classification algorithms and new broad- phase collision detection for real- time sensor data. We demonstrate their performance on the PR2 robot with active sensing. Katja Mombaur and Henning Koch (ORB, IWR, University of Heidelberg): Improving humanoid walking motions by model- based optimization Humanoid robots are extremely complicated dynamical systems for which the generation of whole- body motions, in particular walking is no easy task since the number of parameters to tune for a behavior is very high. Human and humanois are redundant, underactuated systems that require dynamic stabilization. But while human gaits are at the same time efficient, robust and versatile but gaits of humanoids or bipedal robots are at best good in one of these areas. This problem is not only linked to the present hardware, but also to a large extent to the control principles and the software used. The new EU FP7 project KoroiBot Improving humanoid walking capabilities by human- inspired mathematical models, optimization and learning aims to address these issues relying to a big part on model- based optimization. With optimization or more precisely optimal control the redundancy can be turned into a benefit rather than a burden, and the best possibly solution (according to specified criteria) can be found. In the first part of this talk the KoroiBot project is presented. In the second part, we present previous research in cooperation with the LAAS- CNRS in Toulouse on generating optimal walking motions on flat ground and over larger obstacles for the humanoid robot HRP- 2.

3 Ken'ichiro Nagasaka (Business Design and Innovation Laboratory, Sony Corporation, Japan) After QRIO: Toward Physically More Interactive Robots After QRIO, we have been developing a whole- body cooperative force control framework to realize physically more interactive robots. The framework comprises two components: 1) Generalized Inverse Dynamics (GID) that determines joint forces satisfying multiple objectives considering task priorities and various constraints and 2) the Idealized Joint Unit (IJU) that generates accurate torque with assigned apparent inertia and viscosity. On top of them, Generalized Stabilizer (GS) that can cope with non- coplanar cases including flight phase stabilizes the momentum of the system considering the contact constraints. Relevant examples examined on real robots and in the dynamic simulation are shown. Francesco Nori (Fondazione Istituto Italiano di Tecnologia, Genova, Italy) Optimal Whole Body Control of the icub humanoid In this talk we will present recent advances in the whole- body control of the icub humanoid. The icub, initially designed for crawling and manipulating, has been recently upgraded with legs quipped with series elastic actuators. This design has been inherited from the COMAN humanoid which is itself a derivation of the icub platform. This talk will present recent advances with specific focus on locally optimal policies for whole- body coordination in presence of multiple concurrent tasks organized in a prioritized stack. Michiel van de Panne (Department of Computer Science, University of British Columbia, Canada) Developing Locomotion Skills using Policy Search We review several examples of optimization based on policy search methods that have been successfully applied towards a wide variety of simulated motion skills for bipedal and quadrupedal systems. Common themes include the use of compact control representations that contain embedded feedback structures, the use of incremental exploration from good starting points, and the careful structuring of cost functions. We speculate on a number of promising open directions for the optimization of motor skills. Yuval Tassa and Tom Erez (University of Washington, Seattle, USA) Synthesis of Complex Behaviors with Online Trajectory Optimization I will present an online trajectory optimization method and software platform applicable to complex humanoid robots performing challenging tasks in real- time, such as getting up from an arbitrary pose on the ground and recovering from very large disturbances. Our goal at the Todorov Lab is to construct intelligent feedback controllers based on simple cost functions, without manual specification of the details of the behavior. We achieve this through online trajectory optimization, also known as model- predictive control (MPC), using indirect optimal control methods (a.k.a shooting). The idea is to optimize a state- control trajectory integrated up to some time horizon starting at the current state, apply the initial control along the trajectory to the system, and repeat. The previous solution is used to warm- start the optimizer, which often yields near- convergence after a single Newton step. The main challenge in MPC is the need to re- optimize

4 movement trajectories in real- time, which for a humanoid performing a complex task is very computationally demanding. Since most of the time is spent computing the dynamics and their derivatives, a fast physics simulator is prerequisite. To this end we have developed a new physics simulator called "MuJoCo" (Multi- Joint with Contacts). It is a C- based, platform- independent, multi- threaded simulator tailored to control applications. We developed several improvements to the iterative- LQG algorithm that increase its efficiency and robustness. We also developed several models of contact dynamics, which yield different trade- offs between physical realism and speed of simulation. Finally, we developed an interactive environment allowing the user to modify the dynamics model, cost function or algorithm parameters, while interacting with the controlled system in real time. Russ Tedrake (CSAIL, MIT, USA) Polynomial Optimization for Humanoid Robots Some years ago, experiments with passive dynamic walking convinced me that finding efficient algorithms to reason about the nonlinear dynamics of our machines would be the key to turning a lumbering humanoid into a graceful ballerina. For linear systems (and nearly linear systems), these algorithms already exist - many problems of interest for design and analysis can be solved very efficiently using convex optimization. In this talk, I'll describe a set of relatively recent advances using polynomial optimization that are enabling a similar convex- optimization- based approach to nonlinear systems. I will give an overview the theory and algorithms, and demonstrate their application to hard control problems in robotics, including dynamic legged locomotion, humanoids, and robotic birds. Surprisingly, this polynomial (aka algebraic) view of rigid body dynamics also extends naturally to systems with frictional contact - a problem which intuitively feels very discontinuous. Eiichi Yoshida (AIST Tsukuba, Japan) Optimization for Motion Retargeting In this talk we address issues related to optimization for motion retargeting for humanoids. We are exploring a new application of a humanoid robot as an evaluator of assistive device, by using or wearing them instead of human subjects. It is thus important for the humanoid to reproduce motions as close as human. A framework is presented that incorporates robotic motion generation with computer animation through optimization. We also introduce a method that quantifies the differences between a human and a humanoid based on stochastic process.

5