Flow Me if You Can: Uncertainty Is My Co-Pilot
Navigating nano-scale UAVs in forest-like, cluttered environments presents unique challenges. Their compact form factor imposes strict constraints on onboard sensing, power, and compute, limiting the feasibility of traditional perception and control pipelines. To address these limitations, we draw inspiration from biology—specifically the agile, gaze-driven behavior of hummingbirds—to develop an active perception framework that tightly couples motion and sensing. We propose a bio-inspired navigation strategy that uses an uncertainty-aware FlowMotion network to estimate dense optic flow, inter-frame camera motion, and aleatoric flow uncertainty from monocular images. These perceptual cues are used to train a hierarchical reinforcement learning (HRL) agent that actively controls both vehicle motion and camera yaw, enabling agile, perception-driven navigation in complex environments. To bridge the gap between simulation and reality, we introduce VizFlyt, an open-source hardware-in-the-loop platform that leverages 3D Gaussian Splatting for photorealistic rendering at 100 Hz. When deployed on real nano-UAVs within forest-like settings, our trained HRL policy achieves a 96\% success rate in obstacle avoidance and gap-crossing tasks—demonstrating a scalable, compute-efficient, and robust approach to embodied AI navigation in the wild.
Advisor: Professor Nitin Sanket (WPI)
Committee: Professor Jing Xiao (WPI) and Professor Guanrui Li (WPI)
Zoom: Zoom meeting has ended.