Improved Collision Detection through Neural Guided Motion Planning

One of the primary challenges of creating robots that can complete useful real-world tasks is providing the robot
with the ability to navigate through its environment while avoiding collisions with objects. In order to do this,
robotics relies on the use of collision checkers to determine which actions will result in a collision. The time taken
for a robot to come up with a collision-free path depends on how fast it can check for collisions and how often it
has to query the collision checker. We propose to reduce the time spent on collision checking by using data
generated from simulated motion planning problems to train neural networks to predict the best way for a collision
checker to represent objects – decreasing the time taken by a collision detector to determine a collision, as well
as training neural networks to predict how close a proposed path is to objects in the environment – decreasing the
number of times a collision checker needs to be queried.

Faculty Supervisor:

Kamal Gupta

Student:

Partner:

Sanctuary AI

Discipline:

Engineering

Sector:

Information and cultural industries

University:

Simon Fraser University

Program: