Dual object graph and bisimulation metric for object-goal navigation in unfamiliar environment

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Object-goal navigation requires agents to learn visual representations and navigation policies for first-person visual observations in unfamiliar environments.
• Current methods struggle with informative visual representations and robust navigation policies, leading to inefficiencies and deadlocks.

Purpose of the Study:

• To propose novel techniques, dual object graph (DOG) and bisimulation metric (BM), to enhance visual representation learning and navigation policy for object-goal navigation.
• To enable agents to escape deadlock states and improve navigation effectiveness and efficiency in unfamiliar environments.

Main Methods:

• Dual Object Graph (DOG): Integrates current and historical object relationships (category proximity, spatial correlation) to model real-time (COG) and long-term (HOG) object interactions, improving visual representation.
• Bisimulation Metric (BM): A self-supervised reinforcement learning (RL) technique that groups behaviorally similar observations, learning robust latent representations by filtering task-irrelevant information (e.g., background variations, viewpoint changes) to guide navigation policy.
• Both methods are designed to improve navigation policy robustness, helping agents avoid getting stuck or looping.

Main Results:

• Experiments in AI2-Thor and RoboThor environments demonstrate significant improvements in navigation effectiveness and efficiency using the proposed methods.
• The methods enable agents to learn robust latent representations that focus on task-relevant information, aiding navigation.
• Real-world deployment confirmed the transferability and effectiveness of the developed approach.

Conclusions:

• The proposed dual object graph (DOG) and bisimulation metric (BM) effectively enhance both visual representation learning and navigation policy for object-goal navigation tasks.
• These techniques provide a robust solution for agents to navigate unfamiliar environments efficiently and escape common deadlock states.
• The approach shows strong potential for real-world robotic applications requiring autonomous navigation.