Active research projects:
Robust dynamic motion planning, estimation, and control through contact. The ability to control dynamically-complex motions through intermittent frictional contact is a fundamental requirement of any walking, running, and manipulating system. This simple fact has significant computational ramifications, often forcing control system designers to leave smooth optimizations behind and solve more complex mixed-integer or complementarity problems. We are developing new algorithms to design and control dynamic motions through contact while optimizing robustness to errors in contact dynamics (i.e. friction, ordering, timing). External funding: NSF, Sony, and Google.
Nonlinear control for aggressive MAV flight. Autonomous MAVs that operate in challenging flight regimes (e.g., high angle of attack, post-collision) will have tremendous practical impact, from agriculture to disaster response. We are co-designing nonlinear control algorithms and small scale aircraft for achieving high-performance flight in constrained spaces. This project is a collaboration with the Harvard Microrobotics Lab.
Mobile Manipulation. The DARPA Robotics Challenge dramatically demonstrated that some of the world's most advanced robots still minimize contact with their environments and struggle to operate robustly in realistic scenarios. Building upon our work at the DRC, we are developing an autonomous mobile manipulator to attack challenging tasks including automotive maintenance. This project is a collaboration with colleagues at MIT and Draper. External funding: Draper.
|Online adaptive control of soft exosuits. Imagine strapping on a noninvasive exosuit that automatically optimizes itself for your particular gait. In collaboration with Prof. Conor Walsh and the Harvard Biodesign Lab, we are developing efficient learning algorithms for optimizing exosuit controllers online with a human in the loop (Photo credit ). External funding: DARPA.|
Fast optimization algorithms for whole-body control at the DRC. We developed convex optimization-based controllers for stabilizing legged locomotion and implemented them on Atlas, a full-scale hydraulic humanoid robot, for the DARPA Robotics Challenge (DRC). Our DRC teams' software (available here) enabled the Atlas robot to climb ladders, break through walls, open doors, walk over challenging terrain, clear debris, and drive a utility vehicle, all while being commanded remotely over a degraded and delayed network link. External funding: DARPA, ONR.
Variational Bayesian Optimization. Bayesian optimization algorithms are attractive for solving episodic policy search problems on robots, where evaluations are expensive. This work was driven by the observation that 1) a robot's sensitivity to risk, or variance in performance, is not fixed and 2) the variance of cost is often a function of the policy parameters (leading to a heteroskedastic Gaussian process model). We developed Variational Bayesian optimization algorithms and applied them risk-sensitive optimal control problems relating to impact recovery and dynamic manipulation. External funding: NASA.
Learning mobile manipulation skills from demonstration. In collaboration with George Konidaris (Duke), we developed algorithms to perform reliable segmentation of skills, or macro actions, using the uBot-5 mobile manipulator. External funding: NASA.
AAAI'11 Best Student Video: https://www.youtube.com/watch?v=yUICAkSQTZY