Home‎ > ‎

About Critterbot

The Critterbot is an exciting platform for doing research in reinforcement learning and artificial intelligence.

What makes the Critterbot unique?

The Critterbot has a diverse array of sensors it can use to learn about the outside world and its own inner workings.  It can also communicate and interact: it has a small speaker to get people's attention, an array of colorful LEDs, and an elongated tail which it can use to probe and test its environment.

The Critterbot embodies autonomy.  It travels freely, recharges its battery whenever it wishes, and is robust enough to survive dangerous, unsupervised, multi-day trials.  It will help us study new ways of learning computational representations of actions, outcomes, and change.  There is a lot for us to learn from it.

What research is the Critterbot involved in?

A leading problem the Critterbot helps us to solve is the problem of "lifelong learning".  This is the problem of how a software agent can cope with change over the course of its working lifetime, independent of human control.  Our approach to lifelong learning is to task the agent with developing its own computational model of the world, based on questions and experience.

As a simple example of this problem, suppose a human initially supplied the Critterbot with a model of the world: the x and y Cartesian coordinates of its position in a test environment.  This model suffices for simple movement tasks.  But suppose we task the robot with finding the darkest corner in a room.  The best solution to this new problem can change with time, and additional sensors must be incorporated to the world model.

We might be tempted to have a human programmer engineer a new representation for the above task.  But the programmer may not be able to anticipate every task the robot will face in its working lifetime. Tasking the agent with building its own world model is a way for the agent to make productive use of its idle experiences, and its exposure to other tasks it has had experience with.

What are our goals with the Critterbot?

Our broad goals are threefold: to gain experience in connecting a low-level array of sensory inputs to high-level knowledge, to gain experience in teaching and interaction with a lifelike artificial agent, and to evaluate and explore the aforementioned subjective approach to robotics.

In practical terms, our desired outcome is for an agent to be able to build and maintain world models.  These models should allow it to solve complex, unanticipated tasks involving interactions between many heterogeneous sensors.  For instance, we would like to see success in an involved task such as having the robot orient itself into the best position for listening to human conversation, using a world model of its own making.

What is the potential impact of this research?

When we think about infrastructure, we typically think about things like roads, power lines, and waterworks.  One of the costs associated with introducing new infrastructure -- and one which is often forgotten -- is that infrastructure must also be maintained to keep up with changes.

As with any other form of infrastructure, infrastructure that is composed of software (which is becoming increasingly common) also faces the concomitant problem of maintenance.  As we learn about lifelong learning and subjective knowledge representations, we foresee great potential for real-world impact.  In particular, a class of infrastructure with the means to maintain itself over the course of its working lifetime.