research

Some of the research I've done and am doing, categorized for your browsing pleasure

Improving state estimation with data

Robotic perception capabilities in object detection and recognition have grown by leaps and bounds, in part fueled by recent advances in machine learning. However, despite sharing many core fundamentals, state estimation and localization have not seen similar successes. This work explores how machine learning and other data-driven techniques can make state estimation easier and more reliable.

Automatic Parameter Tuning for Robot Perception Systems

2016 - Ongoing

Most perception systems require tuning by a human engineer to adapt to different environments and deployments. Automating this process with reinforcement learning techniques can make it easier and cheaper to deploy perception systems. The primary challenges here are 1. operating without ground truth instrumentation or human supervision to provide feedback on system performance, and 2. efficiently searching a large parameter space to minimize tuning time.

Kernalized Adaptive Kalman Filtering

2016 - Ongoing

The classical adaptive Kalman filter provides a simple and reliable method for calibrating unknown noise covariances on-the-fly by assuming that noise properties evolve smoothly with time. This online estimator can be interpreted as a kernalized estimator, weighting sequences of covariance estimates by their temporal proximity to estimate the current noise covariance. We can extend this idea to kernals operating in more complex feature spaces, for instance utilizing past data that is temporally distant but otherwise similar in context.

Online Calibration of Filter Noise Parameters

2015 - Ongoing

The Kalman filter relies on appropriately-selected noise covariances to perform well. However, selecting these covariances, especially the high-dimensional transition noise covariance, is quite challenging. Instead, we can view the Kalman filter as recursive structure or dynamical belief-space system and use backpropagation to identify both observation and transition covariances.

Fast and Reliable Camera Intrinsics Calibration

2015

We can view camera intrinsics calibration as a statistical model fitting task, allowing us to adapt many useful machine learning concepts. One of particular interest is the idea of covariate shift: training and testing a model on a different distributions of data can result in poor performance. We can apply this concept to intrinsics calibration by using kernel density estimation (KDE) techniques to select a set of calibration images that matches our theoretical test distribution of points.

Predictive Noise Models

2014 - 2015

Standard noise models assume that noise covariance is stationary, or homoscedastic. In some cases, noise may be heteroscedastic and have covariance that varies predictably with factors, for instance radio ranging noise increasing with distance. A stable parametric decomposition of covariance matrices can be used to learn a covariance prediction model, improving filtering consistency and performance.

Environment-independent localization

Localization systems that rely on environmental features can be brittle in unstructured, dynamic deployments, resulting in a tradeoff between localization reliability and acceptable environmental diversity. Alternatively, engineered infrastructure and instrumentation can be used to operate independent of the environment, but is typically costly and difficult to deploy. Below is some of my work on reliable but flexible and inexpensive environment-independent localization.

Multi-Modal Localization for Indoor Mobile Robots

2016 - Ongoing

Perception systems that rely on a single modality can fail catastrophically from single anomalies. For instance, a vision system may be temporarily blinded by sunlight at a particular angle, while a laser rangefinder may not be able to detect infrared-absorbent materials. Combining multiple modes of sensing makes it less likely that any one anomaly will be catastrophic. In this work we combine three odometry systems and two localization systems for highly robust and flexible localization.

Many-Camera Arrays for Mobile Robot Localization

2014 - Present

Calibration, active resource allocation

Fiducial SLAM

2014 - Present

Mapping out large arrays of fiducials manually with survey equipment is impractically laborious and time-consuming. Using a mobile robot to map makes the process much faster.

Active Fiducial Sequences

2014

Fiducial tags have a tradeoff between feature size and the data payload contained in the pattern. Further, they can be heavily affected by lighting, and maintaining large constellations of tags can be time-consuming. An alternative here is to consider active backlit fiducial displays that can be programmatically set to show fast sequences of tags. The tag display is then analogous to a packetized data stream, changing the size-payload tradeoff to a time-payload tradeoff.

Multi-Robot Collaborative Localization

2012 - 2014

Robots working in crowded industrial environments cannot rely on traditional infrastructure-based techniques for localization. Instead we designed a system where robots to sense each other using relative pose information to determine global poses. Sharing this information allows robots to accurately localize relative to each other instead of relying on environmental features.

Unconventional mechanisms

Robotic systems must be flexible and robust, yet powerful, making them at odds with many conventional automation and mechatronic systems. Here is some of my work exploring different ways of building robots.

Pneumatically-driven robotic arms

2017 - Ongoing

Light-seeking behavior on flapping-wing MAV

Summer 2011-2012

Flapping-wing micro aerial vehicles (MAVs) are simple and robust, but suffer from low payloads, restricting the control, computation, and sensing onboard. I worked on adding additional control surfaces and a VGA camera to Stan Baek's I-Bird platform, allowing it to fly towards bright lights.

Google Sites

Report abuse