Drones Learn to Navigate Tight Spaces With AI Control

Drones, or unmanned aerial vehicles (UAVs), are increasingly utilized for diverse applications, from photography to industrial inspections and emergency response. However, their ability to navigate cluttered environments and maneuver through small gaps remains a significant challenge.

To address this limitation, researchers at Zhejiang University have created novel control strategies that enhance drone precision for complex maneuvers, including passage through extremely narrow openings. These strategies, detailed in a Science Robotics publication, employ reinforcement learning (RL) to translate sensor data into motor commands.

The developed sensorimotor policies utilize onboard cameras and internal drone measurements, such as orientation and acceleration, to generate low-level control commands. The RL training process involves a trial-and-error approach, rewarding successful task completion. To overcome exploration challenges in restricted solution spaces, the researchers incorporated an initialization strategy using trajectories generated by a model-based planner.

Both simulations and real-world experiments demonstrated the policies' effectiveness. Drones successfully navigated rectangular and irregularly shaped gaps at various angles, even passing through moving gaps and sequences of closely spaced openings. The system enabled a quadrotor to traverse a 5-centimeter clearance gap at a 90-degree tilt without prior knowledge of the gap's position or orientation.

This advancement has the potential to significantly improve drone agility in challenging, dynamic, and cluttered environments. Future applications could include disaster response for navigating rubble, inspecting confined industrial spaces like pipelines, and exploring subterranean environments such as mines and tunnels.