Title: PerchRL: Enabling Agile, Vision-Driven Perching on Inclined Surfaces During Fast and Unpredictable Movement

Abstract:

Achieving autonomous, vision-guided perching on moving inclined surfaces is a pivotal capability for air-ground coordination, yet it presents significant difficulties due to the restricted field of view (FOV) inherent in such systems. To address this, we introduce PerchRL, a reinforcement learning (RL) framework designed for agile perching on inclined platforms subjected to rapid and irregular dynamics. Our approach utilizes a dual-phase learning protocol, beginning with state-based pre-training and transitioning to vision-based fine-tuning. To enhance generalization across varied platform behaviors, we implement randomized trajectories to mitigate overfitting, alongside temporal augmentation techniques that extract latent motion patterns from historical data. Furthermore, the vision-based fine-tuning phase incorporates a hybrid learning structure featuring visibility-aware state augmentation and active perception rewards, thereby bolstering robustness against intermittent visual obstructions. Comprehensive simulations and physical experiments confirm the feasibility, stability, and real-time efficiency of PerchRL. Additionally, successful implementations on multiple distinct quadrotor models highlight its adaptability. We plan to release the source code to support the broader research community.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC