Search papers, labs, and topics across Lattice.
This paper addresses the challenge of autonomous FPV quadrotor flight in complex environments by utilizing a monocular RGB camera and focusing on the decomposition of optical flow into translational and rotational components. By employing only the translational flow and introducing an uncertainty mask to highlight obstacle structures, the authors enhance the robustness and agility of the flight control policy. Their approach achieves significant improvements, with real-world flight speeds reaching 11.79 m/s and a 93.3% success rate, nearly doubling previous benchmarks for monocular-RGB UAV systems.
By isolating translational optical flow and leveraging an uncertainty mask, this method nearly doubles the flight speed of autonomous UAVs in complex environments.
Autonomous FPV quadrotor flight in complex environments using a monocular RGB camera as the sole exteroceptive sensor remains a fundamental challenge. Recent research has shown that using optical flow as the input of a neural network can achieve end-to-end autonomous flight in cluttered scenes. However, extracting the most relevant information from the flow estimation is the key bottleneck limiting agility and robustness. Existing methods struggle to disentangle obstacle-induced optical flow from the ego-motion background flow and suffer from low signal-to-noise ratios near the focus of expansion (FoE). To address these issues, we decompose the optical flow into translational and rotational components and utilize only the translational flow, which captures scene geometry and depth cues. In addition, we introduce an uncertainty mask derived from inconsistencies between forward and backward flow estimates. This mask highlights obstacle structures, including those within the FoE region. Both cues are fed to a control policy trained in a differentiable simulation framework, which enables efficient first-order optimization across perception and control. We validate our approach through extensive experiments in both simulated and real-world forest environments. The proposed system achieves robust flight at speeds of up to 13.91 m/s in simulation and 11.79 m/s in real-world tests, with a 93.3\% success rate over 30 real-world trials, nearly doubling the previously reported 6 m/s real-world speed of the monocular-RGB optical-flow UAV obstacle avoidance system.
