Workshop: Novel Methods for Learning and Optimization of Control Policies and Trajectories for Robotics
|Katja Mombaur, Gerhard Neumann, Martin Felis, Jan Peters
|Date and Time:
|Friday, May 10, 2013, 8:30 - 18:30
The current challenges defined for robots require them to automatically generate and control a wide range of motions in order to be more flexible and adaptive in uncertain and changing environments. However, anthropomorphic robots with many degrees of freedom are complex dynamical systems. The generation and control of motions for such systems are very demanding tasks. Cost functions appear to be the most succinct way of describing desired behavior without over- specification and appear to underlie human movement generation in pointing/reaching movement as well as locomotion. Common cost functions in robotics include goal achievement, minimization of energy consumption, minimization of time, etc. A myriad of approaches have been suggested to obtain control policies and trajectories that are optimal with respect to such cost function. However, to date, it remains an open question what is the best algorithm for designing or learning optimal control policies and trajectories in robotics would work. The goal of this workshop is to gather researchers working in robot learning with researchers working in optimal control, in order to give an overview of the state of the art and to discuss how both fields could learn from each other and potentially join forces to work on improved motion generation and control methods for the robotics community. Some of the core topics are:
- State of the art methods in model-based optimal control and model predictive control for robotics as well as inverse optimal control
- State of the art methods in robot learning, model learning, imitation learning, reinforcement learning, inverse reinforcement learning, etc .
- Shared open questions in both reinforcement learning and optimal control approaches
- How could methods from optimal control and machine learning be combined?
The workshop will consist of presentations, posters, and panel discussions. Topics to be addressed include, but are not limited to:
- How far can optimal control approaches based on analytical models come?
- When using learned models, will the optimization biases be increased or reduced?
- Can a mix of analytical and learned models help?
- Can a full Bayesian treatment of model errors ensure high performance in general?
- What are the advantages and disadvantages of model-free and model-based approaches?
- How does real-time optimization / model predictive control relate to learning?
- Is it easier to optimize a trajectory or a control policy?
- Which can be represented with fewer parameters?
- Is it easier to optimize a trajectory/control policy directly in parameter space or to first obtain a value function for subsequent backwards steps?
- Is less data needed for learning a model (to be used in optimal control, or model-based reinforcement learning) or for directly learning an optimal control policy from data?
- What applications in robotics are better suited for model-based, model-learning and model-free approaches?
All of these questions are of crucial importance for furthering the state-of-the-art both in optimal control and in robot reinforcement learning. The goal of this workshop is to gather researchers working in robot learning with researchers working in optimal control, in order to give an overview of the state of the art and to discuss how both fields could learn from each other and potentially join forces to work on improved motion generation and control methods for the robotics community.
Pieter Abbeel, U.C.Berkeley, USA
Tim Bretl, Univ. of Illinois, USA
Abderramane Kheddar, CNRS, France and AIST, Tsukuba, Japan
Petar Kormushev, liT Genova, Italy
Richard Longman, Columbia University, New York, USA
Freek Stulp, ENSTA-ParisTech, Paris, France
Konrad Rawlik, University of Edinburgh, Scotland
Oskar von Stryk, TU Darmstadt, Germany
|Gerhard Neumann, Katja Mombaur
|Introduction by the organizers
|A Geometry-Based Approach for Learning from Demonstrations for Manipulation
|UC Berkeley, USA
|Oskar von Stryk
|Optimal control of bio-inspired compliant robots: A trajectory optimization, control or design problem?
|TU Darmstadt, Germany
|10:00 - 10:30
|Guaranteed constraint fulfillment in semi-infinite optimization planning in humanoid robots
|AIST Tsukuba Japan / CNRS France
|Reinforcement and imitation learning of robot motor skills
|Difficulties Specifying Optimization Criteria and Then Making Hardware Perform Model Based Optimized Trajectories
|Columbia University, USA
|Policy Improvement Methods: Between Black-Box Optimization and Episodic Reinforcement Learning
|ENSTA-ParisTech, Paris, France
|12:30 - 14:00
|Information-Theoretic Motor Skill Learning
|Optimal Control: Two Applications in Robotics and Biomechanics
|When is inverse optimal control easy?
|Univ. of Illinois, USA
|15:30 - 16:00
|Probabilistic Inference and Stochastic Optimal Control
|Univ. of Edinburgh, UK
|16:30 - 18:20
|Poster Teasers and Poster Session
Call for Posters
The following submissions have been accepted for poster presentation
- Nonlinear adaptive hybrid control by combining Gaussian process system identification with classical control laws, Jan-P. Calliess, Mike Osborne and Stephen J. Roberts
- Handling High Parameter Dimensionality in Reinforcement Learning with Dynamic Motor Primitives, Adria Colome, Guillem Alenya, and Carme Torras
- Trajectory Optimization for Predictable and Legible Motion, Anca D. Dragan and Siddhartha S. Srinivasa
- Optimizing Robotic Single Legged Locomotion with Reinforcement Learning, Peter Fankhauser, Marco Hutter, Christian Gehring, and Roland Siegwart
- Learning model-based control for lightweight robotic arms, Domagoj Herceg, Dana Kulic, Ivan Petrovic
- SEDS-2: Generating Stable, Reactive, and Robust Robot Motions with Smooth Regression Techniques, S. Mohammad Khansari-Zadeh and Aude Billard
- Trajectory Planning with Adaptive Probabilistic Models, Marin Kobilarov
- Gaussian Belief Space Planning for Articulated Robots, Alex Lee, Sachin Patil, John Schulman, Zoe McCarthy, Jur van den Berg, Ken Goldberg and Pieter Abbeel
- Probabilistic Movement Primitives, Alexandros Paraschos, Gerhard Neumann and Jan Peters
- Learning Musculoskeletal Dynamics with Non-Parametric Models, Katayon Radkhah, Roberto Calandra, and Marc Peter Deisenroth
- Principles for an Alternative Design of Movement Primitives that Uses Probabilistic Inference in Learned Graphical Models, Elmar Rückert, David Kappel , Gerhard Neumann, Marc Toussaint and Wolfgang Maass
- A Geometry-Based Approach for Learning from Demonstrations for Manipulation, John Schulman, Jonathan Ho, Cameron Lee and Pieter Abbeel
- Finding Locally Optimal, Collision-Free Trajectories with Sequential Convex Optimization, John Schulman, Jonathan Ho, Cameron Lee and Pieter Abbeel
March 15 - Deadline of submission for Posters
March 23th - Notification of Poster Acceptance
Extended abstracts (1 pages) will be reviewed by the program committee members on the basis of relevance, significance, and clarity. Accepted contributions will be presented as posters but particularly exciting work may be considered for talks. Submissions should be formatted according to the conference templates and submitted via email to firstname.lastname@example.org
Katja Mombaur , Universitaet Heidelberg
Gerhard Neumann, Technische Universitaet Darmstadt
Martin Felis, Universitaet Heidelberg
Jan Peters , Technische Universitaet Darmstadt and Max Planck Institute for Intelligent Systems
Location and More Information
The most up-to-date information about the workshop can be found on the ICRA 2013 webpage.