Baxter helps us investigate and demonstrate the advantages of a learning by
demonstration approach in the context of underwater autonomous teleoperation.
We are developing a system that will allow underwater remotely operated vehicles
(ROVs) to perform a wide variety of tasks that require dexterous motions.
Applications range from marine biology and archaeology to structural
monitoring and maintenance. Our vision is to develop the cognitive tools that
will allow robots to autonomously complete useful tasks.
We use Baxter as a working example of a teleoperation system. Baxter's left arm
is used for kinesthetic demonstrations of skills, while Baxter's right arm
reproduces the motions that the learned model generates. Baxter's series elastic
actuation allows us to use one of the arm as a safe teleoperation device, while
the other can interact with the environment.
We experiment with two teleoperation scenarios. The first involves the
manipulation of a hand-turned valve, where we teach Baxter how to close the
valve, starting from different valve configurations. In the second scenario we
teach Baxter how to perform a "hot-stabbing" movement - a common ROV task, often
performed multiple times during a mission. We show Baxter how to perform this
task and have Baxter generate appropriate motions at multiple targets. With our
approach, Baxter is able to quickly learn how to perform the tasks at hand and
autonomously generate and execute desired motions.