Automatic selection of coordinate systems for learning relative and absolute spatial concepts.

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Robot Interaction

Background:

• Robots in homes and offices require adaptive learning of spatial concepts from user speech.
• Understanding human-defined coordinate systems is crucial for robots to interpret spatial terms like 'left' (relative) or 'living room' (absolute).
• Current methods struggle as humans don't explicitly state their coordinate systems.

Purpose of the Study:

• To propose a novel method enabling robots to simultaneously estimate coordinate systems and spatial concepts.
• To address the challenge of robots learning spatial concepts without explicit coordinate system information.
• To enhance robot's ability to understand and navigate human environments.

Main Methods:

• Introduced RASCAM (Robot Adaptive Spatial Concept Acquisition Method), an extension of the ReSCAM+O probabilistic model.
• Incorporated a latent variable to represent the coordinate system for simultaneous learning.
• Enabled estimation of coordinate systems, reference objects, and word-concept relationships.

Main Results:

• The RASCAM method successfully learned both relative and absolute spatial concepts.
• Accurate selection of the correct coordinate system was demonstrated across three experimental setups.
• The method effectively estimated unspecified information: coordinate systems, reference objects, and concept-word mappings.

Conclusions:

• RASCAM allows robots to simultaneously infer coordinate systems and spatial concepts from human utterances.
• This approach significantly improves a robot's capability to flexibly understand and interact within new environments.
• The ability to estimate multiple unknown factors offers a significant advancement in human-robot spatial understanding.