End-to-End Autonomous Exploration with Deep Reinforcement Learning and Intrinsic Motivation.

Area of Science:

• Artificial Intelligence
• Robotics
• Machine Learning

Background:

• Developing artificial intelligence (AI) agents for efficient exploration in visually rich and complex environments remains a significant challenge.
• Traditional methods often struggle with autonomous navigation and learning from raw sensory data.

Purpose of the Study:

• To formulate the exploration problem as a reinforcement learning task guided by intrinsic motivation.
• To develop an AI agent capable of autonomous exploration in complex 3D environments.

Main Methods:

• The study employs a reinforcement learning framework with curiosity-driven intrinsic motivation.
• Intrinsic motivation is calculated using episode memory, distributed via count-based methods and temporal distance synchronously.
• The approach was tested in 3D maze-like environments.

Main Results:

• The AI agent successfully learned exploration abilities from raw sensory input.
• The agent achieved autonomous exploration across various mazes.
• The learned policy demonstrated robustness against stochastic objects.

Conclusions:

• The proposed intrinsic motivation framework enables efficient and autonomous exploration for AI agents.
• The method shows promise for applications requiring agents to navigate and learn in complex, visually rich environments.