# Reactive Collision Avoidance Using Real-Time Local Gaussian Mixture Model Maps

In unknown, cluttered environments, robots require online real-time
mapping and collision checking in order to navigate robustly. Discrete
map representations are inefficient for collision checking as they are
expensive in terms of memory and compute. This paper takes a
probabilistic approach to local mapping by representing the
environment as a Gaussian Mixture Model, and leverages geometric
properties to enable efficient collision checking given a
time-parameterized trajectory. In contrast to current
discretization-based methods, a Gaussian mixture model guarantees
probabilistic completeness in its representation and thus preserves
geometric coverage of the environment without losing representation
accuracy with varying map resolutions. We introduce a novel GMM local
mapping algorithm that can be used with a single depth camera
processed on a single CPU, and provide algorithms for collision
avoidance given arbitrary trajectory representations. Finally, we
provide experimentation results demonstrating safety, efficiency, and
map fidelity for real-time collision avoidance with a quadrotor
navigating in a cluttered environment.

## People

M.S., 2020. Now at Shield AI.