Active inference and robot control: a case study.

Area of Science:

• Computational Neuroscience
• Robotics
• Artificial Intelligence

Background:

• Active inference is a theoretical framework for understanding perception and action, rooted in computational and systems neuroscience.
• While prominent in neuroscience, its application and understanding outside these fields, particularly in robotics, remain limited.

Purpose of the Study:

• To demonstrate a proof-of-principle implementation of active inference for controlling a simulated robot arm.
• To investigate the impact of sensory noise on the accuracy of active inference-based robot control.
• To analyze the internal dynamics of active inference to understand multimodal control and sensory integration.

Main Methods:

• Implementation of the active inference scheme for a 7-Degrees-of-Freedom (7-DoF) PR2 robot arm simulation.
• Systematic manipulation of visual and proprioceptive noise levels to assess control performance.
• Analysis of internal system dynamics, including hidden state trajectories, to elucidate framework properties.

Main Results:

• Demonstrated accurate robot arm control under the active inference scheme, contingent on specific noise levels.
• Revealed the multimodal nature of control, highlighting the distinct contributions of proprioception and vision.
• Internal dynamics analysis provided insights into the framework's mechanisms.

Conclusions:

• Active inference offers a viable framework for robot control, with performance modulated by sensory noise.
• The study underscores the potential of active inference for modeling complex phenomena like sensory attenuation and gain control failures.
• Active inference shifts the computational burden from solving inverse problems to constructing generative models, offering a distinct approach compared to optimal control.