CogRob-CoRo Joint Workshop
full-day workshop proposal
25 May 2018 , Brisbane
Room #: P1 (Plaza level)
New horizons in cognitive robotics and AI:
exploiting recent advances for predictive control and prospective interaction between agents
Organizers: TC-CoRo Chairs (Shingo Shimoda, Laurel Riek, David Vernon and Giulio Sandini)
and CogRob Chairs (David Rajaratnam , Maurice Pagnucco)
Abstract
Recent advances in artificial intelligence (AI) have resulted in a wide range of research, commercial and industrial applications. Coupled with developments in sensing technologies, AI has also seen important applications in the field of robotics; especially in areas such as robot vision, mapping and localization.
However, despite these advances, the impact of AI to the area of robot behavior and control has been more modest. While AI based languages and tools for high-level robot control have been developed, nevertheless there is still considerable effort required to understand how to robustly embed these high-level control paradigms within low-level robot controllers and sensors.
Robots are generally controlled with feedback control loops. Feedback control loops respond to environmental changes in real-time, coping with the uncertainty of the environment through controller robustness. One of the important characteristic of a strong coupling of AI and robot control is to control robot behaviors with predictions of future events beyond simple real-time responses to environmental changes.
In this multi-disciplinary workshop, we will discuss the problem of integrating high-level and low-level robot control. The workshop represents a collaboration between the AI-based cognitive robotics community (CogRob) and members of the IEEE Technical Committee on Cognitive Robotics (CoRo). First, we will clarify the problems of merging AI and robot control with invited the speakers from CogRob and RAS TC on Cognitive Robotics. Then we will discuss the importance of higher cognitive functions from both an application and industrial point of view, with respect to the recently developed notions of robot safety and partnership.
Tentative Schedule
Time |
Speaker |
Title |
9:15 – 9:30 |
|
Opening Remarks |
9:30 – 10:00 |
Prof. Giulio Sandini |
Humanizing Robots |
10:00-10:30 |
Dr. Masoumeh Mansouri |
Hybrid Reasoning in Multi-robot Planning |
Coffee break |
||
11:00 – 11:30 |
Prof. Laurel Riek |
Long Term Robot Learning: Modeling and Adaptability |
11:30 – 12:00 |
Prof. Claude Sammut |
Multistrategy Learning for Cognitive Robots |
12:00 – 12:30 |
Dr. Shingo Shimoda |
Robot Intelligence by Tacit learning |
Lunch Break |
||
13:30 – 14:10 |
Short talks for poster session |
|
14:10 – 15:00 |
Poster Session |
|
Coffee break |
||
15:30 – 16:00 |
Prof. Takamitsu Matsubara |
Sample-Efficient Reinforcement Learning for Real-World Robot Control |
16:00 – 16:30 |
Dr. David Rajaratnam |
Cognitive Hierarchies: A Principled Approach to Cognitive Robotics |
16:30 – 17:00 |
Closing Discussions |
|
Presentation Details
Prof. Giulio Sandini
Istituto Italiano di Tecnologia and University of Genova, Italy
Title: Humanizing Robots
Abstract: The performance of humanoid robots has been steadily increasing and nowadays we can claim that sensing and motion abilities of robots are approaching those of humans. This has created the impression that a society where humans and robots co-exist and collaborate is not very far away. Is this true?
During the talk I will argue that robots interacting with humans in everyday situations, even if motorically and sensorially very skilled and extremely clever in action execution are still very much primitive in their ability to understand actions executed by others and that this is the major obstacle for the advancement of social robotics. I will argue that the reason why this is happening is rooted in our limited knowledge about ourselves and the way we interact socially. I will also argue that robotics can serve a very crucial role by joining forces with the communities studying the cognitive aspects of social interaction and by co-designing robots able to establish a mutual communication channel with the human partner (the distinctive mark of human social interaction).
Dr. Masoumeh Mansouri
Örebro University, Sweden
Title: Hybrid Reasoning in Multi-robot Planning
Abstract: My talk is primarily related to hybrid reasoning in Robotics, with a particular focus on multi-robot planning. I argue that GOFAI (“Good Old-Fashioned AI”) methods are not alone sufficient in real-world multi-robot applications. The shortcomings of these methods suggest
that the nuances of real-world robotics problems require combining heterogeneous representations and reasoning (“hybrid” KR&R). I discuss two different approaches to hybrid reasoning for multi-robot planning, namely, general-purpose versus ad-hoc approaches. I summarize the talk with lessoned learned from applying both approaches in an industrial mining application of multi-robot planning.
Prof. Laurel Riek
UC San Diego, USA
Title: Long Term Robot Learning: Modeling and Adaptability
Abstract: In order to build robots that can work longitudinally alongside people,
it is important they are adaptive to people and their environments, and personalize their behavior on the fly. This requires a high degree of personalized, long-term preference learning, which incorporates a dynamic understanding of human context, activities, and goals. This talk will discuss our recent work in this area, within the novel application space of robot-assisted occupational therapy.
Prof. Claude Sammut
University of New South Wales, Australia
Title: Multistrategy Learning for Cognitive Robots
Abstract: Probabilistic and deep learning methods have made substantial contributions to robotics in recent years, but they often suffer from several problems. They usually require many training examples and are difficult to understand. Furthermore, it is difficult to make use of background knowledge, when it is available. Classical AI techniques, using symbolic representations, still have a part to play as they do not suffer the same problems. However, they have their own problems, especially in not dealing, as well as probabilistic methods, with the uncertainty the is inherent in real applications of robotics. We describe a family of hybrid robot software architecture, that include inductive logic programming and planning systems in their decision-making levels, directing reinforcement learning and probabilistic planners in the control and perception layers.
Dr. Shingo Shimoda
RIKEN, Center of Brain Science
Title: Robot Intelligence by tacit learning
Abstract: The capability of adapting to unknown environmental situations is one of the most important factors for the intelligent behavior of the robots. Even in the unexperienced environment and situations, we hope the robots to perform as we planned to overcome the unknow factors by themselves. Tacit learning is bio-mimetic behavior adaptation architecture to tune the roughly-defined robot behaviors to sophisticated ones through body-environment interactions. The robot behaviors created by tacit learning is adapted to the environment in term of motion performance, efficiency and so on. These experimental results suggest the importance of real-time tuning using the bottom-up signals to adapt the behaviors to unknow factors. In this talk, I will introduce the several experimental results to encourage the discussion of appropriate combination of top-down and bottom-up learnings.
Prof. Takamitsu Matsubara
Nara Institute of Science and Technology
Title: Sample-Efficient Reinforcement Learning for Real-World Robot Control
Abstract: Reinforcement learning has been applied in a broad range of robot control scenarios, however, its application to real-world robots still remains difficult since a prohibitively long-time experiment for collecting sufficient data samples is often required. Therefore, developing sample-efficient reinforcement learning algorithms is of primary importance. In this talk, we introduce two reinforcement learning algorithms, Kernel Dynamic Policy Programming (KDPP) and Deep DPP (DDDP) as our recent progress along with this direction. Then, we share our successful applications of RL for 1) robotic cloth manipulation and 2) assistive strategy design in exoskeleton.
Dr. David Rajaratnam
University of New South Wales, Australia
Title: Cognitive Hierarchies: A Principled Approach to Cognitive Robotics
Abstract: Many modern robotic applications call for robots that can exhibit adaptive behaviours and operate in unstructured environments. Such a robot must be able to perform abstract reasoning about its environment, its goals, and the goals of other agents. Importantly, such high-level reasoning needs to occur with respect to the sensing, motion, and computational capabilities of the robot itself. To date, many approaches to building such complex robotic systems have been largely ad-hoc; using software frameworks to plug together components that implement algorithms for vision, mapping, navigation, manipulation, and reasoning. The result is that, while the capabilities of the individual components may be well understood, the system as a whole is poor specified and understood, often resulting in brittle and unpredictable behaviour.
In this talk I discuss a formal framework for tackling this problem. The framework consists of an abstract meta-theory that describes nodes in a cognitive hierarchy, with each node encoding a sub-system that maintains its own belief-state and generates behaviour. Nodes are formalised abstractly, with the representational details of individual nodes remaining opaque at the framework level, thus allowing for the integration of different representations and reasoning mechanisms. I argue that such a principled approach is a crucial if we are to build adaptive, complex robots that can operate reliably and robustly in unstructured environments.
Poster Session Short talk
13:35 – 13:40 | Dr. Klaus Raizer |
Mutual trust in collaborative cognitive robotics | |
13:40 – 13:45 | Mr. Hiroki Kogami |
Muscle Activity Analysis of Physical Therapist Intervention during Standing-up Motion of Hemiplegic Patients | |
13:45 – 13:50 | Dr. Fady Alnajjar |
Intelligent Agent to assess and assist attention and engagement level of children with ASD | |
13:50 – 13:55 | Dr. Alvaro Costa |
Lower Limb Exoskeleton Pattern Control Based on the Attention Level Measurement from EEG Signals Oriented to Rehabilitation | |
13:55 – 14:00 | Dr. Naoki Akai |
Towards Social Autonomous Navigation Utilizing ML Techniques | |
14:00 – 14:05 | Mr. Shotaro Okajima |
What is key factor for robot control with symbolized behavior purpose? | |
14:05 – 14:10 | Dr. Erik Sjoberg |
Virtual Environments for Dynamic Prediction of Interacting Agents | |
Contact Organizers
Dr. Shingo Shimoda
RIKEN Brain Science Institute
shimoda[at]brain.riken.jp
Prof. David Rajaratnam
The University of New South Wales
David.Rajaratnam[at]unsw.edu.au