Probabilistic Programming for Robotics

Autonomous agents use noisy sensors and actuators to interact with a complex external world. As such, inference engines and point estimators are essential to making modern robots work. Often, these are highly specialized and optimized to run in real-time on autonomous agents. These methods are hand-coded and typically make approximations to gain speed, e.g linearization, approximate independence, or static world assumptions. Extended Kalman filters (EKF) are used in a variety of applications from low-level state estimator for feedback control, to high-level mapping applications.

As a high-level language and inference tool, we believe that probabilistic programming has much to offer to the robotics community. As robots become increasingly capable, they also pose greater risks. A wrong state estimate can result in serious damage and injury when robots are equipped with powerful motors. Eliminating the possibility of coding errors in the inference engine by compiling them from high-level specifications would be beneficial both from a safety and design-time perspective. Conversely, robotics can be a rich source of challenging inference problems related to robust long-term autonomy that probabilistic programming might help answer. We present a specific example problem of a common estimation task in autonomous robot and show how PP can address a (vexing) practical calibration issues that practitioners face.

[Extended Abstract][Poster]

This entry was posted in Uncategorized. Bookmark the permalink.