Bidirectional transition consistency between multi-domain observations for visual reinforcement learning...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Visual reinforcement learning (RL) struggles with generalizable representations under visual interferences.
• Existing methods often focus on visual consistency or precise model-based transitions, limiting robustness.

Purpose of the Study:

• To introduce the Multi-Domain Bidirectional Transition (MDBT) model for learning robust and transferable representations in visual RL.
• To address the challenge of policy learning with high-dimensional image observations and diverse visual interferences.

Main Methods:

• MDBT incorporates multi-domain observations with visual perturbations and kernel regions.
• A Data Transformation module diversifies observations, while a Bidirectional Transition module extracts task-relevant dynamics.
• A Consistency Target ensures noise removal and task relevance in learned representations.

Main Results:

• MDBT achieved state-of-the-art performance on the DeepMind Control Suite, improving success rates by 2.3%.
• Significant improvements in generalization were observed on robotic manipulation tasks (up to 124.1%) and CARLA driving simulations (up to 23.2%).
• The model demonstrated enhanced robustness under severe weather and lighting perturbations.

Conclusions:

• MDBT effectively learns robust and transferable representations for visual reinforcement learning.
• The proposed model overcomes limitations of prior approaches in handling visual interferences.
• MDBT shows strong potential for real-world applications requiring reliable visual perception.