HiFly Introduction
Automation is increasingly common in many industrial sectors with the help of robotics. These machines offer a wide array of usages thereby increasing productivity and versatility. With the integration of artificial intelligence, they can interact with the real world and accomplish even more complicated tasks beyond what they were previous pre-programmed to do. The addition of intelligence to robotic automation has many benefits, with the more prominent ones being an increase in efficiency and productivity while maximizing throughput.
Among robotics, unmanned aerial vehicles (UAV), or mini flying robots, have been rapidly growing in popularity. Drone have garnered the attention of developers and hobbyist from the tech industry due to its unique aspect of controllable flight, which differentiates it from other technological innovations. Various business and government entities have already leverage drone capabilities to carry out operations.
Huawei Ascend is a unified platform for AI, offering modern AI solutions to tackle challenges in various industries. As shown in the figure below, it covers the entire industry chain, including the Ascend series of processors, Atlas series of hardware, Compute Architecture for Neural Networks (CANN), AI computing framework, application and development tool chain, administration and maintenance tools, industry applications and services.
Edge computation plays an essential role to reliable and faster data management. With edge computation, we can deliver actionable solutions more quickly using data collected on the drone. The Atlas 200 DK is an edge device from the Atlas family that comes with onboard inference capabilities. Leveraging the Atlas 200 DK's onboard capabilities eliminates the need to perform data transfer to the cloud for inference, reducing computation and time requirements thereby bringing deep learning closer to the drone.
We introduce the HiFly platform for developers and hobbyist interested in the cross-disciplinary of UAV and edge computing. HiFly is a deep learning platform designed for edge-to-drone applications to make ML integration as simple for developers and hobbyists as possible through its modularized capabilities. The platform is designed to run on the Atlas 200 DK powered by the Huawei Ascend ecosystem to bring ML closer to the drone for wider autonomy.
Developers using HiFly can quickly swap out a computer vision algorithm with another during their development process. HiFly enables developers to quickly show a proof-of-concept or build a drone that can achieve complex task within the bounds of ML.
HiFly is still a new platform and there is much work to be done, we are continuously expanding the platform with new ideas. With HiFly, we make it simple to interact with a drone and utilize its capabilities. We look forward to seeing how the community will utilize and advance this platform.
One of the key design goal of HiFly is to enable users to effortlessly integrate off-the-shelf ML models into their applications. HiFly bridges the gap between ML and drone applications through modularized deep learning models. This feature allows users to specify the modular components to be used during their development process as part of their system without having the need to go through model training and evaluation or code modification to fit the pipeline. Users can either train their own model in their preferred framework and then perform model conversion for application development, or simply visit the Ascend ModelZoo for pre-trained models. The Ascend ModelZoo is a model library that is continuously expanded and supported by the Ascend ecosystem backed by an active community of experienced researchers and engineers.
AI-enabled edge-to-drone technology stands at the intersection of robotics, IoT, and AI. We believe it is a promising direction to pursue especially with the recent advancement in such fields. We are actively collaborating with universities to foster innovative solutions from students and we believe HiFly will improve overtime and reach our goal quicker through open-source. We appreciate your contributions and encourage you to provide suggestions.