UAV Collision Avoidance in Unknown Scenarios with Causal Representation Disentanglement DOI Creative Commons
Zhun Fan,

Zihao Xia,

Che Lin

et al.

Drones, Journal Year: 2024, Volume and Issue: 9(1), P. 10 - 10

Published: Dec. 25, 2024

Deep reinforcement learning (DRL) has significantly advanced online path planning for unmanned aerial vehicles (UAVs). Nonetheless, DRL-based methods often encounter reduced performance when dealing with unfamiliar scenarios. This decline is mainly caused by the presence of non-causal and domain-specific elements within visual representations, which negatively impact policies. Present techniques generally depend on predefined augmentation or regularization intended to direct model toward identifying causal domain-invariant components, thereby enhancing model’s ability generalize. However, these manually crafted approaches are intrinsically constrained in their coverage do not consider entire spectrum possible scenarios, resulting less effective new environments. Unlike prior studies, this work prioritizes representation presents a novel method disentanglement. The approach successfully distinguishes between data. By concentrating aspects during policy phase, factors minimized, improving generalizability DRL models. Experimental results demonstrate that our technique achieves reliable navigation collision avoidance unseen surpassing state-of-the-art deep algorithms.

Language: Английский

UAV Collision Avoidance in Unknown Scenarios with Causal Representation Disentanglement DOI Creative Commons
Zhun Fan,

Zihao Xia,

Che Lin

et al.

Drones, Journal Year: 2024, Volume and Issue: 9(1), P. 10 - 10

Published: Dec. 25, 2024

Deep reinforcement learning (DRL) has significantly advanced online path planning for unmanned aerial vehicles (UAVs). Nonetheless, DRL-based methods often encounter reduced performance when dealing with unfamiliar scenarios. This decline is mainly caused by the presence of non-causal and domain-specific elements within visual representations, which negatively impact policies. Present techniques generally depend on predefined augmentation or regularization intended to direct model toward identifying causal domain-invariant components, thereby enhancing model’s ability generalize. However, these manually crafted approaches are intrinsically constrained in their coverage do not consider entire spectrum possible scenarios, resulting less effective new environments. Unlike prior studies, this work prioritizes representation presents a novel method disentanglement. The approach successfully distinguishes between data. By concentrating aspects during policy phase, factors minimized, improving generalizability DRL models. Experimental results demonstrate that our technique achieves reliable navigation collision avoidance unseen surpassing state-of-the-art deep algorithms.

Language: Английский

Citations

0