Active Vision for Robot Manipulators Using the Free Energy Principle.

Area of Science:

• Robotics
• Artificial Intelligence
• Computational Neuroscience

Background:

• Robots face environmental sensing limitations (occlusions, limited field of view/resolution) from single observations.
• Active vision strategies are necessary for robots to gather information through multiple viewpoints for task completion.

Purpose of the Study:

• To frame the problem of active vision as active inference for robotic agents.
• To develop a novel generative model for fusing multi-view data in an object-reaching task.

Main Methods:

• Applied active inference principles, where agents minimize expected free energy using a generative model.
• Developed a deep neural network-based generative model to fuse multiple views into an abstract representation.
• Trained the model by minimizing variational free energy and validated experimentally on a robotic manipulator.

Main Results:

• The active inference approach successfully guided exploratory behavior in a simulated object-reaching task.
• The robot naturally adopted an exploratory strategy, preferring higher vantage points when the target was not in view.
• The end effector accurately reached the target object once it was located through multi-view observation.

Conclusions:

• Active inference provides an effective framework for robots to manage uncertainty and actively explore their environment.
• The proposed deep generative model enables efficient fusion of multi-view data for robotic manipulation tasks.
• This research demonstrates emergent, intelligent exploratory behavior in robots driven by minimizing expected free energy.