Background
I am a Postdoctoral Fellow in the Laboratory for Computational Sensing
and Robotics at Johns Hopkins University, Department of Computer
Science, where I am working jointly
with
Prof. Gregory Hager and
Prof. Noah Cowan.
I received my Ph.D. degree in Computer Science from Rice University in
July 2008. I worked jointly with
Prof. Lydia
Kavraki (Ph.D. advisor) and
Prof. Moshe Vardi.
Research Synopsis
My research goal is to significantly increase the ability of robots to
plan and act on their own or provide assistance in human-machine
cooperative tasks in complex domains. Applications targeted in my
research include
mobile
robotics,
robotic-assisted
surgery,
robot manipulation,
hybrid systems, and
robotic networks.
Synergistic Integration of AI Planning and Motion Planning
-
Novel multi-layered approaches that synergistically integrate AI
planning and motion planning to increase the ability of robots to plan
and act on their own
-
Given a high-level task, these approaches automatically plan the sequence of motions the robot
needs to execute so that the resulting trajectory is dynamically feasible (even when dynamics are
nonlinear), avoids collisions with obstacles, and satisfies the high-level specification
-
These multi-layered approaches make it possible to
plan valid continuous trajectories that satisfy high-level tasks given
by sophisticated mathematical models, such as Finite State Machines,
Linear Temporal Logic, and Planning Domain Definition Languages.
Mobile Robotics
-
Applications on challenging motion-planning
problems with high-dimensional dynamical models of ground and aerial
vehicles demonstrate speedups of orders of magnitude over related
work. These improvements also make it possible to effectively
incorporate physics-based simulations into planning,
which further increase realism in the planned motions by modeling
friction, gravity, and other interactions of the robot with the
environment.
Medical Robotics: Robotic Minimally Invasive Surgery (RMIS)
-
Language of Surgery: Creating statistical models to express RMIS procedures as sentences in a
structured language. Models are derived from analysis of motion and video data gathered during
RMIS performed by expert surgeons on the daVinci robotic system
-
Human-Machine Cooperation or Automatic Surgical Task
Performance:
Novel motion-planning methods that leverage the "language of surgery" models to assist surgeons in RMIS and to
enable robotic systems to automatically perform surgical tasks skillfully
Robotic Manipulation through Haptic Sensing, Exploration, and Planning
-
Novel approaches for mapping and perceiving simultaneously to enable robots equipped with haptic
sensors to pick up and manipulate unknown objects, purposefully explore their surfaces, and identify
the objects based on geometric and tactile appearance information gathered during haptic exploration
Hybrid Systems: Automatic Discovery of Safety Violations
-
Hybrid systems go beyond continuous models by employing discrete logic
to instantaneously modify the underlying dynamics and switch to a
different mode in order to respond to mishaps or unanticipated
changes. Hybrid systems provide sophisticated mathematical models
being used in robotics, air-traffic control, and systems biology.
-
Enhanced automation of air-traffic control provides a viable approach
to deal with the rapid increase in air traffic and alleviate the task
of human operators. Guaranteeing aircraft safety necessitates
high-level specifications of conflict-resolution protocols that take
into account traffic conditions. As the complexity of these protocols
increases, safety verification becomes more challenging, surpassing
the capabilities of current methods.
-
The merit of my research lies in
the ability to discover witness trajectories that demonstrate safety
violations. By asking the framework to exhibit examples satisfying the
negation of the safety assertion, it can be possible to verify with
arbitrarily high probability that the safety assertion
holds. Simulations with numerous airplanes demonstrate speedups of
orders of magnitude over related work.
Approximate Nearest Neighbors and Dimensionality Reduction
-
Approaches to efficiently and accurately approximate the nearest-neighbors graph to significantly
reduce the major bottleneck in nonlinear dimensionality reduction while maintaining accuracy
-
Applications to biology reliably extract from molecular simulation data the main nonlinear modes
of motion while reducing the time to analyze the data from several
months to just a few hours
Large-Scale High-Performance Distributed Computing
-
Distributed computation of very large nearest-neighbors graphs with
millions of points with hundreds of dimensions, yielding near linear
speedup over hundreds of processors
-
Distributed motion-planning methods providing a platform to solve
high-dimensional problems with hundreds of dimensions for articulated
or multi-robot systems
Educational and Research Software