Drones can navigate in unexplored environments using liquid neural networks
In a series quadrotor closed loop control experiments, drones were subjected to range tests, stress testing, target rotation and obstruction, hiking with opponents, triangular objects loops, and dynamic tracking. The drones tracked moving targets and performed multi-step loops in environments never before seen, outperforming other cutting-edge competitors.
The team believes the ability to learn and generalize from expert data while understanding a task could make autonomous drone deployments more cost-effective and reliable. They noted that liquid neural networks could be used to enable autonomous air mobility drones for environmental monitoring and package delivery.
Ramin Hasani, MIT CSAIL’s Research Affiliate, says that the experimental setup in our work tests various deep-learning systems on their reasoning abilities in controlled and straightforward situations. There is much more room for research and development in the future on AI systems that can handle complex reasoning problems in autonomous navigation. This has to be tested first before they are safely deployed in society.
Source:
https://news.mit.edu/2023/drones-navigate-unseen-environments-liquid-neural-networks-0419
