Real-time Robot Path Planning Around Complex Obstacle Patterns through Learning and Transferring Options

Presenter Information

Olimpiya SahaFollow

Advisor Information

Prithviraj Dasgupta

Location

UNO Criss Library, Room 225

Presentation Type

Oral Presentation

Start Date

3-3-2017 1:30 PM

End Date

3-3-2017 1:45 PM

Abstract

Extra-terrestrial navigation and path planning has been an area of substantial interest in the robotics community. For completely adopting an autonomous model, improved path planning techniques need to be developed which will invoke reliable navigation in an unfamiliar environment without human intervention and thus avoid communication delays. We consider the problem of robot path planning in an environment where the obstacle details are initially unknown but the robot can reuse relevant knowledge about collision avoidance learned from previous experience. We propose an algorithm that enables robots to improve their path planning capability by dynamically learning new environmental patterns and corresponding maneuvers from their past navigational experience and probabilistically predict the path to be adopted in presence of partial environmental information. We have conducted several simulated experiments and found that our algorithm takes 24% planning time and 39% total time to solve the same navigation problem compared to a recent, sampling-based path planner.

COinS
 
Mar 3rd, 1:30 PM Mar 3rd, 1:45 PM

Real-time Robot Path Planning Around Complex Obstacle Patterns through Learning and Transferring Options

UNO Criss Library, Room 225

Extra-terrestrial navigation and path planning has been an area of substantial interest in the robotics community. For completely adopting an autonomous model, improved path planning techniques need to be developed which will invoke reliable navigation in an unfamiliar environment without human intervention and thus avoid communication delays. We consider the problem of robot path planning in an environment where the obstacle details are initially unknown but the robot can reuse relevant knowledge about collision avoidance learned from previous experience. We propose an algorithm that enables robots to improve their path planning capability by dynamically learning new environmental patterns and corresponding maneuvers from their past navigational experience and probabilistically predict the path to be adopted in presence of partial environmental information. We have conducted several simulated experiments and found that our algorithm takes 24% planning time and 39% total time to solve the same navigation problem compared to a recent, sampling-based path planner.