Workshop @ IROS 2015

ws_title_koroibot_h2r

A joint workshop by the EU FP7  projects KoroiBot and H2R

Organizers:

Katja Mombaur, ORB, IWR, University of Heidelberg  (KoroiBot), kmombaur@uni-hd.de
Diego Torricelli, CSIC, Madrid, Spain (H2R), diego.torricelli@csic.es

Time & Place:

Sept 28, 2015, 9 :00- 18:00
Hamburg, Germany
More information about the IROS Conference: www.iros2015.org

Workshop program:

Click here to see program

pic_ws_iros2015

Objectives

Understanding human walking and teaching humanoid robots to walk in a human-like way is a challenging task in robotics. It is also one of the central goals of the European projects KoroiBot and H2R as well as the DARPA robotics challenge. To improve humanoid walking it is crucial to identify the essential characteristics of human movement and transfer it to robots. Different geometries and inertial properties of the human and humanoid systems including different kinematic and dynamic constraints have to be taken into account in this transfer.  Model-based optimization, reinforcement learning, movement primitives and neural primitives represent different bio-inspired approaches to motion generation for humanoid robots. The aim of this workshop is to present the different advantages of all these approaches as well as many promising works on combining them. Optimization or optimal control can be performed on robot and human models of different complexity taking different constraints into account, in both offline and online context. It is very useful for exploiting the physical limits of a system, but solutions might have to cope with model-reality mismatches which have to be addressed. Reinforcement learning can work without any model, iterating over reality, but in the contact of complex systems and motions that easily fail, it requires good starting data. Different types of movement primitives, such as kinematic or dynamic primitives (in different senses) provide a good approach  to standardize motions taking into account constraints. Neural primitives do not refer to the explicit motion but to the signal processing side of movement.  While there is a tight coupling between movement primitives and learning (although not reinforcement learning), since movement primitives are usually learned from human data, there are not yet many links nor between optimization and learning nor between optimization and movement primitives. In this workshop, state of the art methods in the fields of optimization, learning and primitives are presented, and a special focus is put on different possibilities of combining optimization methods with the other fields.

Topics of interest

  • Optimization / Optimal control methods for motion generation
  • Reinforcement learning
  • Nonlinear model-predictive control / online optimization of motions
  • Movement primitives (different types of movement primitives, segmentation approaches, retargeting)
  • Dynamic and kinematic models of humans and robots
  • Model reduction
  • Model-free vs. model-based learning
  • Combining optimization & learning
  • Combining optimization & movement primitives
  • Considering multibody dynamics and constraints in movement primitives

Confirmed Speakers:

The workshop will include a combination of tandem talks and regular talks. Tandem talks are joint interdisciplinary talks by researchers from two different research fields and groups about ongoing collaborations.

Tandem talks:

  • Philippe Souères, LAAS-CNRS, Toulouse and Albert Mukoskyi, University of Tübingen:
    Learning movement primitives for the humanoid Robot HRP-2
  • Debora Clever, University of Heidelberg and Dominik Endres, University of Marburg:
    Combining optimization and movement promitives to control humanoid robots
  • Ivan Koryakovskiy, TU Delft and Manuel Kudruss, University of Heidelberg: Combining Model predictive control methods and reinforcement learning approaches for bipedal walking
  • Jose Gonzalez, CSIC, Madrid and Massimo Sartori, Göttingen: Control of human locomotion using neuromuscular primitives

Individual Talks:

  • Florentin Wörgötter, University of Göttingen: Neural control of movement
  • Oussama Khatib, Stanford University: SupraPed for Locomotion in 3D Unstructured Environments
  • Karsten Berns, TU Kaiserslautern: Definition of motor skills learning strategies and cost functions
  • Giovanni de Magistris, JRL CNRS-AIST, Tsukuba: Design of optimized flexible soles and walking pattern generators for humanoid robots
  • Vittorio Lippi, University of Freiburg: Learning in the context of DEC control
  • Katja Mombaur, University of Heidelberg: Optimality in human movement

Supporting Technical Committees:

This workshop is supported by the following IEEE RAS Technical Committees

  • TC Model-based optimization for Robotics
  • TC on Humanoid Robots