Fly360: Omnidirectional Obstacle Avoidance within Drone View
| Authors | Xiangkai Zhang et al. |
| Year | 2026 |
| Field | AI / ML |
| arXiv | 2603.06573 |
| Download | |
| Categories | cs.RO, cs.AI |
Abstract
Obstacle avoidance in unmanned aerial vehicles (UAVs), as a fundamental capability, has gained increasing attention with the growing focus on spatial intelligence. However, current obstacle-avoidance methods mainly depend on limited field-of-view sensors and are ill-suited for UAV scenarios which require full-spatial awareness when the movement direction differs from the UAV's heading. This limitation motivates us to explore omnidirectional obstacle avoidance for panoramic drones with full-view perception. We first study an under explored problem setting in which a UAV must generate collision-free motion in environments with obstacles from arbitrary directions, and then construct a benchmark that consists of three representative flight tasks. Based on such settings, we propose Fly360, a two-stage perception-decision pipeline with a fixed random-yaw training strategy. At the perception stage, panoramic RGB observations are input and converted into depth maps as a robust intermediate representation. For the policy network, it is lightweight and used to output body-frame velocity commands from depth inputs. Extensive simulation and real-world experiments demonstrate that Fly360 achieves stable omnidirectional obstacle avoidance and outperforms forward-view baselines across all tasks. Our model is available at https://zxkai.github.io/fly360/
Engineering Breakdown
Plain English
Fly360 addresses a critical gap in drone obstacle avoidance by enabling UAVs to detect and avoid obstacles coming from any direction, not just ahead. Current systems rely on forward-facing cameras with limited field-of-view, which fails when drones need to move sideways or backwards while their heading points elsewhere. The paper introduces a benchmark with three representative flight tasks and proposes Fly360, a solution leveraging panoramic/omnidirectional perception to achieve full-spatial awareness. This enables collision-free motion in complex environments where obstacles can appear from arbitrary angles—a requirement for practical autonomous drone operations in constrained spaces.
Core Technical Contribution
The core novelty is formulating omnidirectional obstacle avoidance as an explicit problem setting where motion direction decouples from heading direction, requiring 360-degree environmental awareness rather than forward-only perception. The authors constructed the first benchmark for this problem with three representative flight tasks, establishing evaluation criteria for panoramic drone scenarios. Fly360 likely combines panoramic image processing (from fisheye or multi-camera systems) with directional motion planning, enabling the drone to generate trajectories that account for obstacles in all directions simultaneously. This shifts the paradigm from narrow field-of-view forward avoidance to full-spherical spatial reasoning, a fundamental architectural change in how UAV perception and control interact.
How It Works
Fly360 operates on panoramic input from 360-degree sensors mounted on the drone, providing complete environmental context unlike traditional forward-facing cameras. The system processes this omnidirectional perception to build a spatial representation of obstacles in all directions around the UAV. The motion planner then decouples the drone's heading (which direction the camera points) from its velocity vector (which direction it actually moves), allowing sideways and backwards flight while maintaining awareness. The planning component generates collision-free trajectories by reasoning about obstacles relative to the intended motion direction rather than the drone's facing direction, likely using a combination of geometric path planning or learned policies that understand 3D occupancy from panoramic input. The output is a command sequence that safely navigates the UAV through complex environments by leveraging 360-degree obstacle information at each timestep.
Production Impact
For drone systems operating in constrained spaces—warehouses, indoor delivery, infrastructure inspection—Fly360 eliminates crashes caused by blind spots behind or to the side of the drone. Engineers would need to retrofit drones with omnidirectional sensors (fisheye cameras, lidar rings, or multi-camera arrays), adding hardware cost and weight. The computational overhead of processing 360-degree imagery and decoupled motion planning is significant; production systems must optimize panoramic feature extraction and ensure planning latency stays under 50-100ms for safe flight. Integration requires retraining end-to-end control policies or adapting motion planning algorithms to work with the new decoupled motion model, likely involving synthetic data generation using the benchmark and domain adaptation techniques. The payoff is substantially safer autonomous flight in GPS-denied, cluttered environments where traditional forward-facing systems would fail.
Limitations and When Not to Use This
Fly360 assumes omnidirectional sensor availability (panoramic cameras or full lidar coverage), which adds cost and latency compared to single forward-facing cameras—this may not justify the added complexity for simple structured environments or high-speed forward flight. The benchmark is likely simulation-based or limited to specific environments; generalization to real-world outdoor conditions with dynamic obstacles, lens distortion artifacts, or varying lighting is untested. The decoupled motion model assumes the drone can accelerate/decelerate and turn in any direction, but real UAVs have dynamics constraints and momentum that may conflict with sharp omnidirectional maneuvers. The paper does not address computational efficiency for embedded drone platforms; panoramic processing on small edge devices with tight power budgets remains an open problem. Additionally, omnidirectional obstacle detection may suffer from occlusion—panoramic systems still cannot see through walls or into blind spots created by large obstacles.
Research Context
This work builds on decades of obstacle avoidance research in robotics and autonomous vehicles, but specifically addresses the underexplored case of omnidirectional UAV motion. Prior work (like optical flow, depth estimation, and reactive navigation) focused on forward-centric perception; Fly360 extends this to full-spherical scenarios relevant to modern drone applications. The paper likely contributes a benchmark that will become standard for evaluating panoramic perception systems in UAV research, similar to how KITTI advanced autonomous driving evaluation. This opens a new research direction: can learned models (vision transformers, neural radiance fields) better exploit panoramic input for 360-degree planning than traditional geometric methods? The work also connects to spatial intelligence trends—the intersection of 3D understanding, embodied AI, and real-time decision-making on flying robots.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::