SARAH TANG

Learning for unlabeled multi-robot coordination

We present a learning approach to goal assignment and trajectory planning for unlabeled robots operating in 2D, obstacle-filled workspaces. More specifically, we tackle the unlabeled multi-robot motion planning problem with motion constraints as a multi-agent reinforcement learning problem with some sparse global reward. In contrast with previous works, our framework is a general approach that is applicable to arbitrary robot models.

Dissertation

In December 2018, I graduated with my PhD in Mechanical Engineering and Applied Mechanics from the UPenn GRASP Lab!

MIQP trajectory planning for quadrotors with cable-suspended payloads

In this work, we propose an optimization-based trajectory planning algorithm that allows for incorporation of large, yet controlled, payload swings when appropriate. We demonstrate two novel capabilities: maneuvering through difficult obstacles, such as windows whose height is smaller than the cable length, and rapid payload pick-up and releases without the quadrotor moving directly above the payload. This is the first experimental demonstration of these types of maneuvers, which could enable automate rapid Christmas tree harvesting and delivery to hard-to-access disaster sites.

Vision-based geometric control of suspended payloads

Geometric controllers can theortically guarantee stable control of the quadrotor-with-suspended-payload system through configurations where the payload is swung far from the vertical. However, this control approach has not yet been realized in practice. In this work, by using data from a downward facing camera and an onboard IMU, we are able to accurately estimate the payload state and utilize a geometric controller to robustly control agile maneuvers that include payload swings of up to 50 degrees from the vertical. This is the first experimental demonstration of closed-loop feedback control in the three-dimensional workspace for this system and represents a step towards automated payload maneuvering in real-world applications.

Automated payload delivery with quadrotor teams

Most multi-robot systems currently used consist of approximately first- or second-order vehicles. Instead, we consider a team of quadrotors, each carrying a cable-suspended payload, and develop an algorithm for safe simultaneous navigation of all payloads to designated goal positions. We do not constrain the payloads to remain vertical, but instead, allow them to swing upwards when it is safe to do so. This work is the first demonstration of a multi-robot team with vehicles of this level of complexity and is applicable to tasks such as construction, where a single crane performing sequential tasks can be replaced by multiple quadrotors.

Scalable multi-robot trajectory generation with QPs

Trajectory planning for multi-robot teams is a difficult problem --- the problem's search space grows exponentially with the number of robots and each robot often contains kinemtic or dynamic constraints that its trajectory must satisfy. In this light, we propose a two-step algorithm for multi-robot planning. First, a fast (though sub-optimal) motion planning algorithm designs local roundabout-like collision avoidance maneuvers to navigate all robots safely to designated goal positions. Then, a trajectory smoothing process generates a safe, dynamically feasible trajectory for each robot. Unlike past works that optimize all robots’ trajectories in a single optimization problem, our algorithm constructs an independent optimization problem for each vehicle, giving it polynomial, instead of exponential, computation time complexity with respect to the number of robots. We show that this method can safely navigate quadrotors in close proximities.

Real-time planning with an onboard RGB-D sensor

Bringing robots out of motion-capture arenas and into real-world environments is a challenging task that requires interfacing planning algorithms with onboard perception and mapping algorithms. In this work, we propose a method for finding safe, dynamically feasible trajectories through an occupancy grid map built using an onboard RGB-D sensor and demonstrate the execution of the complete obstacle detection to mapping to trajectory execution pipeline on a quadrotor.

Market-based task allocation for cooperative mapping with heterogenous ground robots

Deploying robots into unknown regions to create maps is a common application for ground robots. Using a heterogenous team for this task allows for mapping across different types of terrains and data collection from different sensors. In this work, we describe a framework for autonomous exploration using such a team and demonstrate its usage in modeling a rugged tunnel enviornment.

Decentralized control for shark tracking with a team of autonomous underwater robots

Collecting data on the behavior of long migratory fish species is crucial to the field of marine biology. Unfortunately, current methods for fish tracking suffer significant drawbacks. Satellite tags are able to track fish across long distances, however, can only gather measurements when the fish are at the water’s surface. Acoustic tags are able to function underwater, but require researchers to manually track and follow the fish on boats, which is labor intensive and unsustainable over large distances. Autonomous Underwater Vehicles (AUVs) could potentially solve both these challenges. They can be equipped with acoustic receivers for accurate tracking and have long battery lives for continuous data collection. In this work, we develop a control mechanism that allows a team of AUVs to follow and track a leopard shark.

Benchmarking 3D descriptors for shape retrieval

3D shape retrieval is the process of identifying the subset of 3D models in a collection that are the same as a given query. This problem is challenging, as models can be in different poses or represented from different viewing angles. Many works have focused on developing shape descriptors, scalar or vector quantities calculated at chosen points on the meshes, to determine two models’ similarity. However, proposed descriptors are often evaluated independently and on different datasets. This paper systematically compares a number of prominent 3D shape descriptors and benchmarks their performance.

UAV Learning Workshop at ICRA 2019

I co-organized a workshop on Algorithms and Architectures for Learning-in-the-Loop Systems" at ICRA 2019. Thank you to everyone who attended, especially our wonderful invited speakers!

Research talks at Google Brain and Nvidia

Thank you to Google and Nvidia for allowing me to share my research, "Dynamic manipulation with aerial robot teams"!

University of Pennsylvania iTalks

In 2015, I participated in Penn’s iTalks competition, an event where participants gave 12-minute TED-talk style presentations that were judged by a faculty panel and the student audience on clarity of presentation, impact of work, and interdisciplinary nature of research. This talk won the Best Presentation (determined by faculty judges) and the Audience Favorite (determined by audience voting) Awards.

Princeton University Center for Information Technology Policy panel

"This talk is the second in our “Can Law Keep Up with New Technology?” series of lunch timers. Each program explores the current state of an emerging technology and the legal and ethical considerations that stem from it. Peter Asaro and Sarah Tang will discuss non-military drones: what is possible now and in the near future using drone technology and how we should think about their effect on privacy in public space, considering surveillance and remote sensing capabilities."

Tutorial talk at the Technology Management Conference at Saab Dynamics

Drones learn tricks with suspended loads, through-window package delivery inevitable

Penn’s GRASP Lab receives $5.5 million for ‘fast, light and autonomous’ flying robots

Survey of advancements towards fast flight on quadrotors

Survey of networked robotics

Aerial Robotics Coursera course

Our Coursera course on Aerial Robotics, part of a 5-course series on Robotics from the Penn GRASP Lab, is now live!

GRASP Research Experience for Undergraduates

During the summer of 2014, I worked with an incredible REU student on a research project titled “Autonomous Flight and Landing of a Quadrotor on a Moving Ground Vehicle Using April Tag Vision-Based Control”, which won the Best Presentation Award!

Army Educational Outreach Program, Research & Engineering Apprenticeship Program

During the summer of 2017, I served as a mentor for the Army Educational Outreach Program, Research & Engineering Apprenticeship Program. I worked with a high school student to quantify and analyze inter-robot aerodynamic effects in a team of bitcraze Crazyflie robots during formation flight.

USA Science and Engineering Festival

In April 2016, the GRASP lab hosted a booth at the USA Science and Engineering Festival, featuring our running, hopping, rolling, and flying robots!

FIRST Lego League

From 2014–2016, I was honored to volunteer as a judge for the FIRST Lego League Championships at Penn!

Rube Goldberg Celebration

From 2011–2012, I served as the Science Programming Assistant for the Cotsen Children’s Library at Princeton University, where I helped organize and facilitate events at the intersection of science and literature. One of our most successful initiatives was the first Rube Goldberg celebration, which featured this book-turning Rube Goldberg Machine (constructed with Tanner DeVoe).

Princeton Engineering Education for Kids (PEEK)

I was involved in Princeton Engineering Education for Kids (PEEK) from 2009–2013 as a volunteer and 2010–2012 as co-project coordinator. Our flagship initiative was a five-lesson series with every third grade classroom in the Princeton Regional School District. We also worked with a middle school engineering club, the Princeton Public Library, and presented at one-time outreach events throughout the school year.