With an escalating demand for quadcopters in emergency response, inspection and delivery, the need for speed and energy efficiency is paramount. However, developing autonomous systems for aggressive high-speed flight faces challenges, particularly in creating computationally efficient optimal control algorithms.
Recent advancements in the field have seen the adoption of neural network controllers trained through supervised or reinforcement learning. Yet, a reality gap emerges during the transfer from simulation to real-world scenarios, necessitating the use of robust inner-loop controllers during actual flights. This is a limiting factor to the network’s control authority and overall flight performance.
In the recent article in Robotics and Autonomous Systems, we introduced G&CNet which adopts a novel approach by enabling quadcopters to operate without depending on inner-loop controllers for stabilization, providing direct motor commands instead. This approach gives full control authority to the network, but it is more sensitive to modeling errors. To mitigate this challenge we train an “adaptive” G&CNet that learns to compensate for unmodeled moments. This adaptive G&CNet can identify and optimize motor commands in real-time, even in the presence of significant disturbances like external weight additions. Performing this onboard with a fraction of the computational load is crucial for the development of smaller and lighter optimal flying robots.
2024 |
End-to-end neural network based optimal quadcopter control Journal Article In: Robotics and Autonomous Systems, vol. 172, 2024, ISSN: 0921-8890, (Funding Information: This work was supported by the European Space Agency.This research was co-funded under the Discovery programme of, and funded by, the European Space Agency. Funding Information: This work was supported by the European Space Agency . Publisher Copyright: © 2023 The Authors). |