Featherlight Flyers: Mastering Cluttered Environments with Minimalistic Sensing and Reinforcement Learning

Shenqi Wang, Stein Stroobants, Christophe De Wagter, Guido De Croon

Keywords: Reinforcement learning; Obstacle avoidance; Edge AI

Abstract: Flapping-wing Micro Aerial Vehicles (FWMAVs) offer high agility and bio-inspired flight, enabling applications in environments where conventional quadrotors struggle. Reinforcement learning (RL) has recently shown great promise for collision-free autonomous navigation, yet the strict payload and computational limitations of FWMAVs make traditional approaches based on cameras or heavy onboard computation impractical. In this work, we propose a minimalistic obstacle-avoidance framework for FWMAVs that leverages lightweight multizone time-of-flight (ToF) sensors combined with recurrent Proximal Policy Optimization (PPO). We introduce a bio-inspired state estimation model that infers velocity and position from body attitude dynamics without odometry. Experiments in simulation demonstrate that our RL-based approach outperforms state-of-the-art handcrafted ToF-based baselines and surpasses vision-based methods in some scenarios. Furthermore, we validate the method in real-world experiments with a Flapper Nimble+ robot.

Submission Type: Novel research

Student Paper: Yes

Demo Or Video: No

Public Extended Abstract: No (Only the title, author list, and abstract will be released on OpenReview)

Submission Number: 10