Integrating high-level and low-level cognitive capabilities is essential for developing robotic systems that can adaptively act in our daily environment in active collaboration with humans.
The main objective of this workshop is to share knowledge about the state-of-the-art machine learning methods that contribute to modeling high-level cognitive capabilities in robotics, and to exchange views among cutting-edge robotics researchers who are interested in adaptive high-level cognitive capabilities in robotics.
Our daily environment is full of uncertainty with complex objects and challenging tasks. A robot is not only required to deal with things appropriately in a physical manner but also required to perform logical and/or linguistic tasks in the real world.
Consider when a human user asks a robot ``please move it into the blue box.'' The robot has not only to solve a manipulation task, moving the target object to a particular blue box but also to estimate what ``it'' represents. In addition to solving the manipulation task, the robot should estimate the meaning of ``into'' representing the relationship between ``it'' and ``the blue box'' in the real-world environment. When a robot attempts to communicate and collaborate with human users in a real-world environment, e.g., the RoboCup@Home environment, bridging high-level and low-level cognitive capabilities appropriately is critical. The high-level cognitive capabilities include logical inference, planning, and language. In contrast, the low-level cognitive capability includes physical control, behavioral motion generation and sensory perception.
Conventionally, symbol-based and/or rule-based approaches have been employed to model high-level cognitive capabilities in robotics. However, it has been pointed out that such conventional methods could not create a robot that can deal with the uncertainty that is inevitably found in the physical environment and natural human-robot communication.
However, recent advances in machine learning techniques, including deep learning and hierarchical Bayesian modeling, are providing us with new possibilities to integrate high-level and low-level cognitive capabilities in robotics. It became clearer and clearer that machine learning methods are indispensable to create a robot that can deal with uncertainty.
In this workshop, we will investigate how to create synergies so that advanced learning of sensorimotor and cognitive capabilities can interact to create a bootstrapping effect in different levels of skill acquisition. We believe the topic of this workshop is timely and necessary for the IEEE-RAS community.
The workshop consists of the following parts.
Introduction by the organizers
Deep learning and Bayesian nonparametrics for robotics
Machine learning for processing natural language and speech signals
Machine learning for high-level cognitive models
We call for poster presentations publically. Selected papers will be shown in a spotlight talk session.