IJCAI-2001 Workshop on

Reasoning with Uncertainty in Robotics

Seattle, Washington, August 4-5, 2001
[ Workshop Home | Call for Papers | Organizers | Final Program ]

Invited Talks

SAIL robot for Autonomous Mental Development under Uncertainty
By Juyang Weng (Michigan State University, USA)

How could a robot handle the tremendous uncertainty in multimodal sensory streams such as vision, speech, and touch in order to derive its internal representation and to produce context-dependent, complex behaviors? The uncertainty arises from not only noise and variation in the sensory input, but also a lack of knowledge about the related context at any moment in a multimodal sensing environment. We approach this fundamental issue through autonomous online mental development, motivated by human mental development from infancy throughout the life span. We describe the SAIL-3 developmental program, which enables the SAIL robot to autonomously develop its skills through online, real time interactions with the physical environment using its sensors and effectors. The sensors include video cameras for vision, microphones for audition, and force sensors for touch. The effectors include pan-tilt head for visual attention, arm and gripper for object manipulation and driving wheels for navigation. During the development, human teachers are considered as a part of the environment, interacting with the SAIL robot online through its sensors. We present the SAIL-3 mechanisms for mental development and their working. A basic engine called Hierarchical Discriminant Regression (HDR) incrementally generates mental architecture and representation for online memory self-organization and updating. As a general purpose regressor with a logarithmic time complexity, it is compared with neural networks, supporting vector machine (SVM), and other tree classifiers such as CART and C5.0. Three modes of learning are integrated by the SAIL-3 developmental program: supervised learning, reinforcement learning and the new communicative learning.

In the communicative learning, language acquisition and teaching using language are carried out in the same autonomous mode. We show how the SAIL robot learns simple spoken instructions interactively and how humans use spoken commands to interactively teach the robot to manipulate objects and to perform vision-guided navigation.

Classifying and Explaining Robotic Uncertainty: An Illustrated Catalog
By Enrique H. Ruspini (SRI International, USA)

Any intelligent agent operating in complex, unstructured, environments must make decisions relevant to its performance on the basis of knowledge that it is less than ideal. Intelligent mobile autonomous agents (robots) are, because of inherent limitations in their perceptual abilities, a foremost example of cognitive entities that must cope with uncertainty about their own state and that of their operational environment.

In this talk, we discuss the nature of uncertainty, in general, and its relevance to various problems found in the design of intelligent autonomous agents, in particular. We present first, in the context of a basic unifying model, basic epistemic notions of uncertainty, evidence, and ignorance, focusing on the explanation of probabilistic and non-probabilistic approaches to measure and characterize uncertainty.

We discuss next several situations and tasks where mobile autonomous agents must cope with probabilistic and non-probabilistic uncertainty. We focus particularly on non-probabilistic approaches to uncertain reasoning and explain the connections between this family of methods with logics of utility and with measures of uncertainty like the radius of information of Traub and Wozniakowski.

We present several examples of application of techniques inspired on these ideas to planning and control of individual robots and of teams of collaborating mobile agents. We employ these examples to discuss the interrelation of communication, perception, motion, and other actions in the performance of robotic tasks.

Identity Uncertainty
By Stuart Russel (University of California Berkeley, USA)

We are often uncertain about the identity of objects. This phenomenon appears in theories of object persistence in early childhood; in the well-known Morning Star/Evening Star example; in tracking and data association systems for radar; in security systems based on personal identification; in the content analysis of web pages; and in many aspects of our everyday lives. The talk will describe a formal probabilistic approach to reasoning about identity under uncertainty in the framework of first-order logics of probability, with application to wide-area freeway traffic monitoring.


Website hosted by AASS
Last updated on July 18, 2001