Object-Centric Scene Representations Using Active Inference.

Area of Science:

• Robotics
• Artificial Intelligence
• Neuroscience

Background:

• Robots require scene understanding to interact with environments.
• Current methods struggle with inferring object properties from raw sensory data.

Purpose of the Study:

• To develop a novel approach for robot scene understanding using an object-centric generative model.
• To introduce a new benchmark for evaluating active vision agents in 3D object viewpoint tasks.

Main Methods:

• Utilized an object-centric generative model for scene understanding.
• Employed active inference, a neuro-inspired framework, for perception and action.
• Developed a benchmark for active vision agents to find optimal object viewpoints in a 3D workspace.

Main Results:

• The active inference agent successfully inferred object category and pose in an allocentric frame.
• The agent demonstrated a balance between epistemic foraging and goal-driven behavior.
• Achieved over double the success rate compared to supervised and reinforcement learning baselines.

Conclusions:

• The proposed object-centric generative model enhances robot scene understanding capabilities.
• Active inference provides an effective framework for robot perception and action.
• The new benchmark facilitates rigorous evaluation of active vision agents.