(Self-organizing Autonomous
Incremental Learner)
SAIL is the latest innovation under development in the Michigan
State University PRIP (Pattern
Recognition and Image Processing)
Lab. SAIL is very different from other existing
sensor-based experimental systems. In developing SAIL, we do not model knowledge-level
rules. Instead, we model how the system can learn knowledge autonomously
and incrementally using its sensors and actuators. A human teacher will
encourage or discourage certain behaviors of the system by issuing score
signals to the system. The system is designed to self-organize the sensory
information and actuator experience learned to maximize the expected score.
Although the concept of machine learning has been investigated extensively
in the Artificial Intelligence (AI) community, the learning systems there
require preprocessed symbolic or numerical data. The methods are not applicable
to learning directly from high-dimensional sensory and actuator data. Furthermore,
the subject of autonomous learning has not received much attention in the
AI community.
The goal of SAIL requires that it has sensors and actuators. The prototype
we are building includes a head that could speak, listen, and see, an arm,
a movable base, and a ``memory'' system that performs functions like those
of the human brain. Apparently, it is a major undertaking in hardware, in
software, and more so in theory and methodology.
SAIL is based on SHOSLIF (Self-organizing
Hierarchical Optimal Subspace
Learning and Inference Framework),
another product of the Michigan State University
PRIP Lab. The difference is, with SHOSLIF,
learning is done in a spoon-feeding fashion and with SAIL, learning is autonomous.
The SAIL robot will recieve information in one or more of the following
ways:
- from microphones which are placed so that it can realize from which
direction the sound originated.
- from two video cameras which are located so that it can calculate distances
to objects.
- from bumpers located on the mobile base unit
- from other various inputs such as motor stalls, etc.
The SAIL robot will then process this input and respond in one or more
of the following ways:
- by moving (using the base)
- by moving the eyes via the pan-tilt units
- by speaking via speakers
- by manipulating the arm
Here is a graphical representation of the way in which these functions
all tie together:
For more information on SAIL, see Laura Blackwood's paper at /user/prip4003/blackwoo/papers/sail2.ps
or Dr. John Weng's paper at /user/prip4003/blackwoo/papers/weng.ps
which can be reached from any of the Computer Science UNIX machines or remotely
by connecting to the cps.msu.edu domain.