Inferring Categories to Accelerate the Learning of New Classes
Proceedings of the IEEE/RSJ International Conference on Intelligent Robots
and Systems (IROS), 2013
Robert Goeddel and Edwin Olson
Abstract -
On-the-fly learning systems are necessary for the deployment of general purpose
robots. New training examples for such systems are often supplied by mentor
interactions. Due to the cost of acquiring such examples, it is desirable to
reduce the number of necessary interactions. Transfer learning has been shown
to improve classification results for classes with small numbers of training
examples by pooling knowledge from related classes. Standard practice in these
works is to assume that the relationship between the transfer target and
related classes is already known.
In this work, we explore how previously learned categories, or related groupings of classes, can be used to transfer knowledge to novel classes without explicitly known relationships to them. We demonstrate an algorithm for determining the category membership of a novel class, focusing on the difficult case when few training examples are available. We show that classifiers trained via this method outperform classifiers optimized to learn the novel class individually when evaluated on both synthetic and real-world datasets.
Abstract -
Despite evidence that human wayfinders consider directions involving landmarks
or topological descriptions easier to follow, the majority of commerical
direction-planning services and GPS navigation units plan routes based on
metrically or temporally shortest paths, ignoring this potentially valuable
information. We propose a methodo for generating directions that maximizes
the probability of a human arriving at the correct destination, taking into
account a model of their ability to follow topological, metrical, and
landmark-based directions. We discuss optimization techniques for employing
these models and present a method, DART, for extracting model-improved sets
of directions in a tractable amount of time. DART employs particle simulation
techniques to maximize the probability that the modeled wayfinder will
successfully reach their destination. Our synthetic evaluation shows that DART
produces improvements in arrival rates over existing methods and illustrates
how DART’s directions reflect properties of the wayfinder model.
Abstract -
Tasks like search-and-rescue and urban reconnaissance benefit from large numbers
of robots working together, but high levels of autonomy are needed in order to
reduce operator requirements to practical levels. Reducing the reliance of such
systems on human operators presents a number of technical challenges including
automatic task allocation, global state and map estimation, robot perception,
path planning, communications, and human-robot interfaces. This paper describes
our 14-robot team, designed to perform urban reconnaissance missions, that won
the MAGIC 2010 competition. This paper describes a variety of autonomous systems
which require minimal human effort to control a large number of autonomously
exploring robots. Maintaining a consistent global map, essential for autonomous
planning and for giving humans situational awareness, required the development
of fast loop-closing, map optimization, and communications algorithms. Key to
our approach was a decoupled centralized planning architecture that allowed
individual robots to execute tasks myopically, but whose behavior was
coordinated centrally. In this paper, we will describe technical contributions
throughout our system that played a significant role in the performance of our
system. We will also present results from our system both from the competition
and from subsequent quantitative evaluations, pointing out areas in which the
system performed well and where interesting research problems remain.