Research

Research
My goal is to explore the fundamentals of robotic mobility and manipulation and neural control. One of the key motivations is to improve the robustness of robots by drawing inspiration from the biological control systems like the human body. How can we build artificial systems that are as reliable as the human body and can assist society? To achieve this, we need to understand the mechanisms in the human body, and, in particular, the human hand which enables our daily activities. However, several open questions about the functioning of the human control system remain. For example, how can we do complex tasks like twirling a pen with such ease, when in contrast robots find it difficult to move beyond pick and place operations? How do the intrinsic properties of the musculo-skeletal system contribute to our abilities? How do we learn new tasks? How can human skill improve robots? Here are some recent accomplishments toward my goal.
Framework for Manipulation with Neuro-musculo-skeletal Redundancy The human hand is an intricate and redundant mechanism with over twenty five muscles that actuate the many joints. We can use electromyography to study muscle use, but we still need a framework for understanding, say, how much force each muscle applies and how that muscle contributes to a task. Furthermore, task definition is itself difficult. Using a task definition that includes end-effector force, stiffness, and posture, we have developed a framework that helps us interpret biological control strategies in a physical context. Check out our IEEE ICRA 2008 and ICRA 2009 papers. Work with Pedram Afshar and Yoky Matsuoka.
Human-Robot Interaction Recently, we have begun exciting work to understand how a human should interface with a high DOF robotic arm and hand for telemanipulaton and prosthetics. What should the division of control hierarchy be? Also, can we harness human skill to improve the performance of robotic manipulation. Work with Yoky Matsuoka and Josh Smith (Intel Research Seattle).
System Identification of the Anatomically Correct Testbed Robotic Hand Our group has built an anatomically correct robotic hand to study the details of the human hand's musculoskeletal system. To develop new control strategies, we did a system identification of the robotic hand. Specifically, we computed the mechanical advantage that each muscle has on each joint, which turns out to be a 6 x 4 matrix of functions. Check out our IEEE BioRob 2008 paper. We hope to have a journal paper on this soon. Work with Ashish Deshpande, Ralph Lin, Brian Dellon, and Yoky Matsuoka.
Motor Learning How does the brain learn new tasks? For example, how does a tennis player learn the coupled motion between the shoulder, elbow, and wrist joints to accurately position the racquet head? It has been shown that the learning in its early stage is energetically expensive and error prone. With more practice, the brain builds internal models of the task, reducing metabolic cost and error simultaneously. However, our recent work has shown that in tasks with little sensory feedback the control system prioritizes task performance over minimizing energy consumption. Specifically, the control system chooses to achieve the required performance level first at high energy cost, and then utilizes internal models to explore low energy solutions. Check out our work in the Transactions of Biomedical Engineering (TBME 2008). Work with Robert Howe and Yoky Matsuoka.
Understanding the Role of Small Actuators in Manipulation The control redundancy of the human hand is largely because each muscle actuates multiple joints, a feature captured by a highly coupled moment arm matrix. Thus, we would like to know to which tasks the intrinsic muscles (in the hand itself) contribute and to which tasks the extrinsic muscles (in the forearm) contribute. We have noticed that even though the intrinsic muscles have low moment arms (mechanical advantage) in flexion tasks, they are used heavily, possibly as low gain control knobs that provide stability and precision. This work has been submitted to ICRA 2009. Work with Yoky Matsuoka.
Novel Control Strategies to Improve Robot Robustness When a robot's conventional operation mode is ineffective, the robot's operation fails. For example, when car's wheel slips in snow, its mobility is compromised. Similarly, when contact fails between a manipulator and an object, the robot loses grip. Such events are called locomotion or manipulation errors. My doctoral work at CMU explored the structure of locomotion errors and devised new ways to approach the mobile-robot recovery problem, including a careful dynamics and kinematics analysis and AI techniques like dynamic programming. Check out my PhD thesis. Work with Matt Mason and Al Rizzi.
Legless Locomotion We have obviously heard of legged locomotion, heck, we ourselves are bipedal. So what is legless locomotion? It is a new locomotion mode my doctoral work discovered. Consider a flipped-over turtle which cannot walk. While it can probably flip back and continue walking, that is not the point. Our work showed that the turtle, or a round-bodied robot, can "walk" by merely swinging its legs smartly---with the right phase, frequency, amplitude, and offset. We call this new oscillatory locomotion mode legless locomotion. The beauty of legless locomotion is that it is an altogether unique locomotion mode, different from all previously studied locomotion modes, and hence spawns some very interesting dynamics, kinematics, and control research topics. This work won the nomination for a Best Student Paper prize at IEEE ICRA 2004. Check out the many publications on this topic: IJRR 2008, IROS 2006, ICRA 2004, ICRA 2004 video, IROS 2004, IROS 2003. Work with Matt Mason and Al Rizzi.
Other Robotics Work Earlier, I have done projects related to RoboCup robotic soccer, fuzzy algorithms, and leg-wheel hybrid robots.