Biomimicking UAV Policy Training For Obstacle Avoidance From Optical Flow With One Camera

The research for unmanned aerial vehicles (UAV) that can navigate through narrow spaces like caves, rubble, crop fields, or factories has resulted in the need for research on the design of insect-sized UAVs. Once developed, the insect-sized UAVs will need fully autonomous navigation systems to avoid obstacles. The inefficiency of stereo vision at the small scale requires researchers to find a way to train a machine learning policy that can avoid obstacles by using a single camera. An obstacle avoidance policy that utilizes optical flow data computed from a camera was first experimented with by Hu et al. (2025). They trained such a policy, through gradient optimization, on a differentiable physics simulation of a simplified UAV model, to then be transferred to a real UAV. The resultant policy fails to tackle the inherent weaknesses of optical flow: zero optical flow around the direction of motion and noise due to rotation. We challenged this problem through two elements of biomimicry: a wide field of view (280-degree fish-eye camera) and view direction controlled by the policy (an extra output value for rotation around the vertical axis). The visual data input was not limited to any direction, so the policy learned to follow safer paths that allow more exploration. Early collision results, edging on 50% (when going with an average speed of 4 m/s, adjusted for a 20 cm in diameter UAV) for an unknown double-layered obstacle course, show the potential to meet and exceed state-of-the-art success benchmarks.