Past Projects:

Mobile Manipulation. With colleagues from MIT and Harvard University, we developed an autonomous mobile manipulator to tackle challenging tasks in the real world (e.g. tasks like automotive maintenance, unstructured warehouse assembly, and automated food prep). We built an integrated architecture with novel modules spanning perception, planning, and manipulation, published in various venues, and was one of five teams in the world to be named a Finalist for the 2017 KUKA Innovation Award.

Featured Publications: [arXiv’17] [IROS’17]

*KUKA Innovation Award Finalist: Real-world Interaction Challenge Finalist

Deep Closed-loop Motion Primitives. We developed an approach to simplify visuomotor skill acquisition for robots using deep learning with closed-loop motor units at the output layer of the network. An interesting byproduct is possible self-supervision derived from a dynamical systems approach capturing the robot’s evaluation of its own actions through closed-loop control interactions with the environment.

Featured Publications: [RSS-WS’16]

Active Belief Recognition. We present an active, model-based recognition system that applies information theoretic measures in a belief-driven framework to recognize objects using the history of visual and manual interactions to select informative interactions. We define belief state over populations of learned affordance models that account for visual transition and design a system that focuses on the impact of the belief-space and object model representations on recognition efficiency and performance.

Featured Publications: [IROS’16] [ICDL’16]

Phase Lag Bounded Velocity Planning. We proposed an algorithm for planning longitudinal vehicle velocities constrained by the bandwidth of the input forcing function (the path curvature) and guarantee bounds on the driving frequency in the heading controller so that the robot minimizes path deviation to the intended path. The result allows the robot to use the full performance envelope of the drive motors and provides a principled means of regulating precision and time performance.

Featured Publications: [Humanoids’15]

FUN3D Physics-based Knowledge Bots. Millions of complex physics-based simulations, usually performed by highly trained and skilled analysts with decades of experience, are required for the design of an aerospace vehicle. We introduced methods to train knowledge bots to identify the idiosyncrasies of simulations and encapsulate such expertise. These algorithms are used to identify regions of instability without any additional information about mathematical form or applied discretization approaches.

Featured Publications: [NASA/TM’16] [NASA/TM’14]

Image result for umass stroke ubotRobot Mediated Stroke Rehabilitation. We presented a novel system design for the extended virtual presence of robot-mediated rehabilitation, providing new analytic tools to further aid therapists in conducting therapy. The integrated system consists of collecting numerous quantitative performance parameters during therapy for analytics and feature representations for adaptive therapy learning.

Featured Publications: [ICORR’15] [ICORR’13]