The study of neural solutions to motor control is the most interesting way forward for robotics. Conversely, robots that reflect the characteristics of biological systems provide a unique opportunity to test theories about neural systems.
I use reinforcement learning, dimensionality-reducing synergies, motor primitives, multiple dynamics models, and motor babbling to control robots. I use EMG, vibrotactile feedback, and muscle models to understand sensing and motor control and to explore how humans and other animals learn to use tools and their own bodies.
|EMG Control, Gestures, and Vibrotactile Sensory Substitution
Incorporating augmentative vibrotactile feedback can improve performance of a virtual object
manipulation task using finger movement. Vibrotactile sensory substitution for prosthetic applications, however, will necessarily not involve actual finger movement for control. I study the utility of such feedback when using myoelectric (EMG) signals for control, on unimpaired participants and upper-limb amputees.
||Dynamic Robotic Manipulation
Given constraints, robots can be made to accomplish particular
tasks; even thousands of years ago Hero of Alexandria described a
machine which could pour wine into glasses. How is it that our robots
remain almost as single-purpose while we humans can do this and then
drink from those glasses without crushing them, perhaps while
absentmindedly writing with a pen? This ability to maneuver arbitrary
objects in arbitrary ways, somehow compensating for the uncertain
surfaces, frictions, and weights of the hand and environment, is
Control of dynamic movements and contact is still an open issue. I am adapting state-of-the-art control methods to achieve
manipulation tasks using a unique robot which mimics the human hand.
These efforts have the potential to improve robotics, but conversely
allow us to answer concrete questions about what must be happening in
brains and bodies.
||Neural Control of Movement
Understanding what enables human-level dexterity is a task which
continually seems within reach, but escapes at the last minute. Controlling our bodies to
manipulate objects in the environment often feels effortless, but we
are beginning to understand that there are elegant biomechanical and
neural resources being used outside of our awareness.
For instance, it has been suggested that the brain employs ”muscle synergies” to make planning, control, and learning more tractable. While the details (even the existence of) of neural synergies are still highly controversial, a robotic control paradigm inspired by that work can reduce dimensionality and address control constraints while providing the richness needed to learn manipulation tasks.
Humans employ context-specific internal models
for their bodies and the world around them. They learn them
through embodiment in the environment, by gathering data in
a task-specific way. Similarly, robots operating under humanlike
task constraints must have the ability to accommodate
situations unknown to their designers.
When the dynamics of the robot (or body!) are nonlinear, and interacting nonlinearly with the environment, for example during contact, a linear approximation to the dynamics may be made. There has been great interest in analytically creating local linear approximations, but this fails when the true nonlinear model is unknown or otherwise poorly estimated. I am interested in using embodied emperical experience of robot performance to create locally valid linear approximate dynamics models which may be combined or switched among depending on context.