Multiple Self-Supervised Auxiliary Tasks for Target-Driven Visual Navigation Using Deep Reinforcement Learning.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Deep reinforcement learning for visual navigation demands extensive environmental interaction and training time due to sparse rewards.
• Current methods often rely on pre-designed features or demonstrations, limiting adaptability and sample efficiency.

Purpose of the Study:

• To enhance sample efficiency and navigation performance in deep reinforcement learning for visual navigation.
• To develop a framework that learns navigation strategies directly from high-dimensional images without manual feature engineering or prior demonstrations.

Main Methods:

• Proposed a novel framework utilizing multiple self-supervised auxiliary tasks for visual navigation.
• Implemented an LSTM-based dynamics model and an attention-based image-reconstruction model as auxiliary tasks.
• Focused on latent representation learning directly from raw, high-dimensional images, bypassing the need for traditional ResNet features.

Main Results:

• The proposed self-supervised auxiliary tasks significantly improved training efficiency.
• The framework demonstrated superior navigation performance compared to baseline algorithms.
• Effectiveness was validated on both simulated and real-world image datasets.

Conclusions:

• The developed framework enhances sample efficiency and navigation performance for deep reinforcement learning.
• Self-supervised learning from raw images offers a viable alternative to feature engineering and demonstrations.
• This approach reduces the reliance on extensive training time and computational resources for visual navigation tasks.