Leveraging Predictions of Task-Related Latents for Interactive Visual Navigation.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Interactive Visual Navigation (IVN) tasks require embodied agents to learn environmental interactions for goal achievement.
• Current reinforcement learning (RL) methods struggle with IVN due to inefficient environment representation in partially observable settings.

Purpose of the Study:

• To introduce a flexible self-supervised RL framework, Predictions of Task-Related Latents (PTRLs), to improve IVN task performance.
• To enhance the learning of structured environment dynamics and sequential observation representations.

Main Methods:

• PTRL framework utilizes self-supervised learning to predict agent's next pose based on actions.
• Incorporates an attention and memory module for spatiotemporal dependency exploitation.
• Employs a state value boost module for adaptation to novel environments via input perturbations and value function regularization.

Main Results:

• PTRLs demonstrate superior performance in increasing accuracy and generalization capacity for IVN tasks.
• Sample efficiency in RL network training is significantly enhanced through modular training and hierarchical decomposition.

Conclusions:

• The proposed PTRLs framework offers a robust solution for improving embodied agent navigation in complex environments.
• PTRLs effectively addresses limitations of traditional RL methods in learning environment representations for IVN.