Autonomous Navigation of drones and robots in indoor and outdoor cluttered environments
Abstract
In this project, we will investigate the use of deep learning and in particular deep reinforcement learning for autonomous navigation of robots and drones in cluttered environments. In particular, we build on top of three existing projects in PI s lab at UC Berkeley. The first one has to do with autonomous navigation of hexapods in cluttered environments. Hexapods are considerably more stable and agile than 2 or 4 legged robots and as such have the potential to maneuver, climb over, tuck under, or get around objects to carry out sensitive missions. We have demonstrated the application of deep reinforcement learning to teach low-cost hexapods to climb over rows of joists in attic like environments. We will extend this work to more general autonomous exploration and navigation in indoor cluttered environments with staircases, large objects and overhanging objects such as chairs and tables. The second project has to do with obstacle avoidance and autonomous navigation of drones in cluttered environments such as forests. Here, we will apply classical techniques to generate trajectories which are then used as supervisory signal in a reinforcement learning framework for the drone with limited sensors to avoid obstacles while approaching the desired target. The third project has to do with drone-based 3D neural surface reconstruction of indoor environments using monocular RGB images captured with a low-cost drone. Our ultimate goal here is to remove the human in the loop for data capture and have the drone autonomously navigate complex indoor environments including staircases and tunnels.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 05, 2025
- Source ID
- FA95502410074
Entities
People
- Avideh Zakhor
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of California Regents