A simultaneous navigation and 3D reconstruction framework for autonomously guiding micro aerial vehicles

Background

The field of autonomous robotics increasingly relies on micro aerial vehicles (MAVs) to navigate complex, unknown indoor environments. Weighing under 100 grams, these agile drones operate in tight spaces inaccessible to larger robots. Consequently, there is a critical need for robust and safe navigation systems tailored to these miniature platforms. However, small MAVs possess severe payload and power limitations, making them physically incapable of carrying traditional, heavy depth-sensing equipment like LiDAR or stereo cameras. Therefore, achieving reliable spatial awareness requires extracting high-resolution metric depth information using only minimal, lightweight sensors, such as a single monocular camera.

Current monocular navigation approaches suffer from significant drawbacks. Many state-of-the-art methods rely on end-to-end learning, mapping camera images directly to control actions. This methodology frequently overfits to specific training environments or robot dynamics, severely limiting generalization to novel spaces. Furthermore, these systems lack explicit metric spatial grounding and 3D reconstruction, making it difficult to accurately gauge obstacle distances. Consequently, they exhibit high collision rates. Existing systems also lack crucial self-arresting mechanisms; without the ability to explicitly recognize when no safe trajectory exists, these drones cannot automatically halt, inevitably leading to crashes in highly constrained or unmapped spaces.

Technology

The disclosed solution is a simultaneous navigation and 3D reconstruction framework for micro aerial vehicles utilizing a single monocular camera and offboard computation. The architecture applies lightweight image transformations to mitigate domain shift before feeding visual data into a pre-trained depth estimation model. This generates metric depth maps that are sequentially fused using a truncated signed distance function to construct a local 3D voxel map. A planning module then evaluates pre-computed motion primitives against this map to dynamically select a collision-free trajectory. Furthermore, the system features a self-arresting mechanism that automatically halts the vehicle if no safe path is identified.

This technology is differentiated by enabling autonomous metric navigation on severely payload-constrained hardware without requiring heavy LiDAR or stereo sensors. Unlike existing end-to-end learning approaches that map images directly to control actions and lack spatial grounding, this solution relies on explicit 3D reconstruction. This spatial awareness allows for superior collision avoidance and enables the unique self-stopping capability. Consequently, the framework achieves a fourfold reduction in collision rates compared to state-of-the-art monocular navigation systems. By combining off-the-shelf fusion methods with pre-trained depth models, it provides a robust, safe, and generalizable navigation stack for complex indoor environments.

Benefits

• Enables autonomous 3D navigation for ultra-lightweight, payload-constrained micro aerial vehicles (MAVs) without requiring heavy or power-hungry sensors like LiDAR or stereo cameras.
• Significantly reduces collision rates compared to existing monocular navigation systems through explicit 3D metric reconstruction and dynamic obstacle avoidance.
• Enhances operational safety with an automatic self-arresting mechanism that halts and lands the vehicle in highly constrained spaces when no safe trajectory is available.
• Improves generalization across different environments and prevents overfitting by utilizing lightweight image transformations to mitigate domain shift for pre-trained depth estimation models.
• Facilitates efficient, real-time indoor navigation by leveraging offboard computation and a library of pre-computed motion primitives to minimize onboard processing requirements.

Applications

• Autonomous warehouse inventory tracking
• Indoor industrial equipment inspection
• Confined space search and rescue
• Autonomous HVAC ductwork inspection
• Covert indoor security surveillance

Additional info

This framework enables micro aerial vehicles to perform autonomous navigation and 3D reconstruction using a monocular camera. Transformer-based models extract metric depth maps, fused into a truncated signed distance function (TSDF) voxel map. A planning module evaluates pre-computed motion primitives against this map to select collision-free trajectories. By incorporating a self-arresting mechanism, the system achieves safe metric navigation on payload-constrained hardware without requiring heavy LiDAR or stereo sensors.