NCREE Study: Safe Hierarchical Reinforcement Learning for MAV Indoor 3D Reconstruction
Smart operations
A study published by NCREE proposes a safe hierarchical reinforcement learning framework for autonomous indoor 3D reconstruction using monocular MAVs, combining control barrier functions and SLAM-aware rewards.
A study published in the NCREE Newsletter presents a safe hierarchical reinforcement learning framework for autonomous indoor 3D reconstruction using monocular micro aerial vehicles (MAVs). The framework decouples long-horizon viewpoint planning from short-horizon navigation, using control barrier functions (CBF) as a real-time safety layer and SLAM-aware feedback to improve tracking stability without true pose data.
Key takeaways
- The architecture comprises high-level viewpoint planning, low-level velocity commands, and a CBF safety filter enforcing collision constraints.
- Experimental results show SLAM-aware feedback improved task success from 80% to 88% and reduced tracking failure rate from 10% to 7%.
Source：國家地震工程研究中心｜地震工程研究簡訊 Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human. Translation checked by AI.