SpatialSim: Recognizing Spatial Configurations of Objects With Graph Neural Networks

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Embodied AI requires geometrical reasoning for goal achievement.
• Identifying and discriminating object configurations is crucial for AI agents.
• Deep learning literature has overlooked this area of spatial reasoning.

Purpose of the Study:

• Introduce SpatialSim, a novel diagnostic dataset for geometrical reasoning.
• Evaluate AI models' ability to perform spatial identification and discrimination tasks.
• Benchmark progress in developing principled AI approaches to spatial understanding.

Main Methods:

• Developed SpatialSim dataset with 'Identification' and 'Discrimination' tasks.
• Tested fully-connected message-passing Graph Neural Networks (MPGNNs).
• Compared MPGNNs against Deep Sets, Multi-Layer Perceptrons, and Convolutional Neural Networks (CNNs).

Main Results:

• MPGNNs demonstrated effectiveness due to relational inductive biases.
• MPGNNs outperformed Deep Sets and Multi-Layer Perceptrons.
• High-capacity CNNs failed on the challenging 'Discrimination' task.
• Current GNNs show limitations on both SpatialSim tasks.

Conclusions:

• Relational inductive biases are key for geometrical reasoning in AI.
• SpatialSim serves as a valuable benchmark for advancing AI spatial cognition.
• Further research is needed to overcome GNN limitations in complex spatial tasks.