LiDAR-based UAV perception is limited by narrow field of view and payload constraints. Traditional fixed-speed scanning lacks scene awareness, degrading performance in complex environments.
AEOS combines model predictive control (MPC) and reinforcement learning in a hybrid architecture inspired by owl sensing behavior. An analytical uncertainty model predicts pose observability while a neural network learns implicit cost maps for optimal scanning control.
Extensive experiments demonstrate significant odometry improvements compared to baseline methods while maintaining real-time performance for onboard deployment.
Key Features & Contributions
🦉 Bio-inspired Design
Inspired by the active sensing behavior of owls, AEOS adapts LiDAR scanning patterns based on environmental context and task requirements.
🤖 Hybrid MPC + RL
Combines analytical uncertainty models for exploitation with neural networks for exploration, achieving optimal scanning control.
🌍 Real-world Transfer
Point cloud-based simulation environment enables effective sim-to-real transfer across diverse real-world scenarios.
⚡ Real-time Performance
Maintains computational efficiency for onboard deployment while significantly improving odometry accuracy.
Bio-inspired Active Sensing System
Owl-inspired Adaptive Scanning: Traditional UAV LiDAR systems use fixed-speed rotations that ignore environmental complexity.
AEOS mimics the selective attention mechanism of owls, dynamically adjusting scanning patterns based on scene structure and uncertainty,
leading to more efficient data collection and improved localization performance in challenging environments.
Active Sensing Mechanical Design
AEOS-DRONE
Mechanical Design
Real-world Experiments
Challenging Environments: AEOS has been extensively tested in complex real-world scenarios including industrial facilities and underground environments.
Our experiments demonstrate significant improvements in odometry accuracy compared to fixed-rate scanning, while maintaining real-time performance constraints
essential for autonomous UAV operations.
Field Testing Results
Industrial Equipment
Basement
Rescue Operation
Learning & Practice in Simulation Environments
Scalable Training Framework: Our point cloud-based simulation environment incorporates real-world LiDAR maps from diverse global locations,
enabling robust policy learning and effective sim-to-real transfer. The lightweight neural network learns implicit cost maps from panoramic depth representations,
balancing exploration and exploitation for optimal scanning strategies across varied environmental conditions.
