Autonomous Scene Exploration for Robotics: A Conditional Random View-Sampling and Evaluation Using a Voxel-Sorting...

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Autonomous robot exploration requires complex skills like scene representation, identifying unexplored regions, and efficient viewpoint selection.
• Existing methods often face challenges in real-time scene mapping and efficient exploration strategy.

Purpose of the Study:

• To propose an autonomous exploration system for robots capable of creating and updating scene representations.
• To estimate the Next Best View (NBV) efficiently for maximizing novel information acquisition.
• To enable robots to explore unknown environments systematically and completely.

Main Methods:

• Utilizes a dual OcTree representation to encode occupied, free, and unknown space within a scene.
• Employs a conditioned random process to generate and evaluate view hypotheses for NBV estimation.
• Incorporates sensor-specific ray-casting and voxel sorting for accelerated assessment of potential views.

Main Results:

• The proposed system successfully explores scenes from scratch, progressively building knowledge.
• Demonstrates efficient exploration by selecting views that yield the most novel information.
• Achieves faster exploration completion times compared to human subjects in experimental scenarios.

Conclusions:

• The dual OcTree-based autonomous exploration system is effective for scene mapping and navigation.
• The NBV estimation strategy significantly enhances exploration efficiency.
• The system offers a promising solution for autonomous robot exploration in unknown environments.