Kevin S. Tracy

I am a PhD student in The Robotics Institute at Carnegie Mellon University advised by Zac Manchester. I am also a student researcher at [Google] Intrinsic, working with Stefan Schaal on contact-rich manipulation.

I did my BS in Mechanical Engineering at Rice University, and my MS in Mechanical Engineering at Stanford University with a concentration in dynamics. At Stanford I served as a TA to Stephen Rock for Control Design Techniques, and Mac Schwager for State Estimation and Filtering for Robotic Perception.

I have also been the head TA for Optimal Control & Reinforcement Learning for the past few years, where my materials and recorded lectures are available on the website.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

I'm interested in differentiable convex and non-convex optimization, motion planning, contact physics, and estimation. I like finding optimization-based solutions to new and old problems.

A robust method for computing smooth derivatives through convex optimization problems solved with a primal-dual interior-point method. The associated Python [JAX] package is able to leverage these smooth derivatives to solve a variety of robotics-inspired problems.

A unified and efficient framework for computing continuous collision detection information between convex sets in a way that allows for smooth differentiation of the collision information with respect to the pose and geometry of each set.

A method for learning contact force models from data in a way that scales very well on a GPU with a novel Linear Model Learning (LML) algorithm.

An ADMM based QP solver that runs on a GPU and can handle very large MPC problems commonly found in robotics.

A quasi-dynamic rigid-body simulator for point clouds with tangential and torsional friction for each of the points. The result is stable and fast simulation with a QP solver.

A framework for computing differentiable collision information betweens a set of convex primitives by solving for the minimum scaling of each shape that results in an intersection.

An updated version of the CPEG algorithm that allows for free-final time trajectories combined with an atmospheric density estimation scheme.

An improved version of dynamic mode decomposition that includes jacobian regularization from an a priori. dynamics model for learned dynamics in optimal control.

The classic square-root Kalman Filter significantly simplified through the use of QR decompositions.

DiffPills computes collision information between capsules and padded polygons by forming and solving differentiable quadratic programs.

CALIPSO is a nonlinear programming solver designed to handle the types of conic and complimentarity problems that arise in robotic motion planning problems with contact.

V-R3x is a three cubesat constellation that demonstrated on-orbit mesh networking and relative navigation.

A control and estimation framework is proposed for a fine pointing cubesat where traditional reaction wheels are swapped for three actuated booms with tip-mounted masses.

CPEG is an online-capable tool for optimal entry guidance in the Martian atmosphere.

The environmental disturbances on a nanosatellite are analyzed within a novel control framework to establish practical performance limits.

A trajectory optimization solver based on differential dynamic programming with an augmented lagrangian to handle conic constraints.

A unified approach to quaternion differential calculus for use in optimization.

A trajectory optimization solver based on differential dynamic programming with an augmented lagrangian to handle conic constraints.

Convex model-predictive control is used to stabilize a flexible spacecraft during planned thruster firings.