Accelerated System Design for Perception and Control in Energy-Constrained UAVs
Abstract
Despite the enormous potential of battery-powered drones in the AF & commercial applications, true autonomy is still not achieved. A pilot constantly navigates the drone using on-board cameras & real-time wireless communication. All components constantly drain battery, & flight time is less than 1 hour. Constant communication makes the drone vulnerable to adversaries. Drone operation & reaction require the onboard compute system to accomplish perception & control. These tasks rely heavily on machine learning & stochastic control algorithms. These algorithms are significantly compute intensive & cannot be accomplished with current general-purpose platforms within the power & payload constraints of battery-powered drones. Without innovative compute systems, such capabilities & true autonomy may remain out of reach, due to the diminishing benefits from traditional transistor scaling & stagnant improvements in general-purpose computing. This project aims to develop a holistic & cross-stack solution–from programming languages down to circuits–that enables perception & control in energy-constrained drones. The key innovation is a manycore accelerator that efficiently executes machine learning & stochastic control algorithms. To support a wide variety of algorithms, we leverage the insight that many of these algorithms are stochastic optimizations that can be effectively expressed using mathematical formulations. Using this, we embark on a basic scientific investigation that develops a domain-specific language for learning & stochastic control; an interpreter that maps the specified algorithms to the instructions of a novel virtual machine; a compiler that schedules & translates the virtual machine instructions to the accelerator reconfiguration; a design of a novel clustered many-core accelerator; operating system primitives that integrate the computing platform; FPGA prototypes; & finally deployment on quadcopters.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2017
- Source ID
- FA95501710274
Entities
People
- Hadi Esmaeilzadeh
Organizations
- Air Force Office of Scientific Research
- Georgia Tech Research Corporation
- United States Air Force