Depth-driven UAV obstacle avoidance with teacher-guided exploration in cluttered environments
TL;DR
SAC-DGPF, a teacher-guided obstacle avoidance framework based on a depth-guided potential field with Soft Actor Critic, reduces overfitting tendencies and avoids performance collapse in experiments.
Abstract
Obstacle avoidance in cluttered environments presents a significant challenge, as it requires rapid perception and decision-making in partially observable three-dimensional space, particularly for unmanned aerial vehicles (UAVs) operating with constrained onboard vision. Compared with traditional map-based approaches, end-to-end methods based on deep reinforcement learning (DRL) directly learn a direct perception to control policy, but they often require large amounts of data and exhibit unstable exploration. To address these issues, we propose SAC-DGPF, a teacher-guided obstacle avoidance framework based on a depth-guided potential field with Soft Actor Critic (SAC). The DGPF module provides safety biased guidance in the early stage of training, while a latent representation learned from depth observations encodes compact spatial and temporal information. This structured state representation helps the policy capture obstacle characteristics and supports a more effective balance between exploration and exploitation during flight. Simulation experiments in cluttered indoor and outdoor scenes show that the proposed method improves sample efficiency and yields consistently higher success rates across randomized start and goal configurations drawn from the same scene. Under training with randomized navigation task configurations, SAC-DGPF reduces overfitting tendencies and avoids performance collapse in our experiments.