Drones stop sorting static from dynamic obstacles: one polar gradient field gives the safe direction, replanning under 4.8 ms

Obstacle avoidance in dynamic environments no longer requires tracking and separating static from dynamic obstacles first; a single safety gradient field directly supplies the avoidance direction. Earlier modular pipelines rely on detection and tracking of moving objects, model dynamic obstacles as bounding boxes or ellipsoids, and apply separate evasion strategies, which can fail in highly cluttered scenes; ESDF-based self-supervised methods must repeatedly build and update distance fields for rapidly changing obstacle configurations, which is computationally expensive. In a 40 m by 40 m cluttered ROS-Gazebo arena with three obstacle configurations, each method flew 25 independent trials per scenario at a maximum speed of 2.5 m/s; in the two dense scenarios with 75% dynamic obstacles at average speeds of 0.5 m/s and 1.0 m/s, the method reached a success rate 16.0% above the best baseline VIGO, while average replanning time stayed under 4.80 ms across all test cases.

The safety gradient field is generated from an ego-centric polar occupancy map aligned with the onboard camera frustum, so it naturally matches the limited visual observation range. The method first projects privileged scene point clouds into an ego-centric polar occupancy representation, then converts it into a differentiable temporal polar gradient field via multi-kernel separable Gaussian convolution, and uses ESDF-guided offline learning to distill explicit geometric priors into the field's structural parameters. The gradient field is computed in parallel on the GPU while the ESDF is built on the CPU; using 2,000 randomly sampled offline frames and measuring 20 iterations after 5 warm-up runs, TPGF consistently took less time than ESDF under batched evaluation.

The framework targets autonomous aerial vehicles that perceive only through an RGB-D camera and continuously plan smooth trajectories toward a goal in unknown dynamic environments; validation covers three obstacle configurations in simulation, three generalization scenarios (static forest, cluttered indoor, crowded corridor), and real quadrotor flights in a cluttered indoor environment and a dynamic forest. The authors propose future work on embedding more robust dynamic obstacle perception into the end-to-end framework, incorporating obstacle and vehicle velocities into the current purely spatial field for real-time adaptive speed adjustment, and extending TPGF to three dimensions.

The authors note that applying convolutions on the polar grid introduces inherent spatial ambiguity, which is why ESDF-guided offline learning is needed to calibrate gradient directions; ablations show that removing the correlation loss or the safe-zone loss degrades the field into a narrow, core-dominated configuration with success rate near zero. In generalization tests, success rate shows an expected degradation trend as obstacle density and flight speed increase together. Real flights cap maximum speed at 1.6 m/s for pedestrian safety.