Beyond Forward Vision: Enabling True Omnidirectional Awareness for Autonomous Drones

Most autonomous drones today suffer from a fundamental sensory limitation. They navigate using forward facing cameras or limited field of view sensors, creating a dangerous blind spot that becomes critical during lateral translation or backward flight. This constraint forces an artificial coupling between the drone's heading and its velocity vector, requiring the vehicle to face the direction of travel to avoid obstacles. In the paper "Fly360: Omnidirectional Obstacle Avoidance within Drone View", Zhang et al. address this limitation by introducing a panoramic vision based framework that enables true omnidirectional obstacle avoidance without explicit 3D reconstruction.

The research identifies a crucial gap in existing literature. While conventional systems assume aligned motion and heading, real world applications such as cinematography, warehouse inspection, and search and rescue frequently require decoupled operation. A drone filming a moving subject must maintain camera orientation while translating sideways through cluttered environments. Current methods fail here because obstacles approaching from the rear or sides remain invisible until collision becomes imminent.

The Architecture: Perception, Depth, and Policy

Fly360 employs a two stage perception decision pipeline that separates visual processing from control generation. At the perception stage, the system ingests 360 degree RGB inputs from dual fisheye cameras and converts these into depth maps using a pretrained panoramic depth estimation model. This intermediate representation proves crucial for bridging the simulation to reality gap. Since training data generation remains computationally cheaper in simplified simulators, using depth rather than raw RGB allows the policy to generalize across visual domains without requiring photorealistic rendering during training.

The decision