Improving Generalization in Collision Avoidance for Multiple Unmanned Aerial Vehicles via Causal Representation...

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Deep reinforcement learning (DRL) shows promise for multi-unmanned aerial vehicle (UAV) collision avoidance and navigation.
• Current DRL methods struggle with generalization, failing to perform well in scenarios outside their training data.
• This limitation is often caused by spurious correlations learned from training data.

Purpose of the Study:

• To address the generalization problem in DRL-based UAV navigation.
• To propose a novel method using causal representation learning to identify robust features.
• To improve the ability of DRL agents to generalize to unseen environments.

Main Methods:

• Developed a causal representation learning framework to extract causal features from images.
• Employed causal intervention to disregard irrelevant factors of variation.
• Integrated these causal representations into the policy network for action prediction.

Main Results:

• The proposed method demonstrated superior generalization capabilities compared to existing state-of-the-art techniques.
• Experimental results showed improved performance across diverse testing scenarios.
• Causal representations effectively mitigated the impact of spurious correlations.

Conclusions:

• Causal representation learning is a viable solution for enhancing DRL generalization in complex tasks like UAV navigation.
• The method offers a pathway to more robust and reliable autonomous systems.
• Future work can explore further applications of causal inference in multi-agent DRL.