Goal-Directed Planning for Habituated Agents by Active Inference Using a Variational Recurrent Neural Network

Area of Science:

• Robotics
• Computational Neuroscience
• Artificial Intelligence

Background:

• Robotic agents often struggle with generalization in action planning due to partial world models derived from sensory-motor experiences.
• Existing forward model frameworks in robotics face challenges with high degrees of freedom, limiting their generalization capabilities.

Purpose of the Study:

• To investigate the efficacy of predictive coding (PC) and active inference (AIF) frameworks for enhancing generalization in robotic agents.
• To develop a generative model that learns a prior distribution in a low-dimensional latent state space for improved action planning.

Main Methods:

• The proposed model employs a generative model within PC and AIF frameworks to learn a prior distribution from habituated sensory-motor trajectories.
• Learning involves inferring optimal latent variables and synaptic weights to maximize the evidence lower bound.
• Goal-directed planning is achieved by inferring latent variables that maximize the estimated lower bound.

Main Results:

• The model demonstrated sufficient generalization in both simple and complex robotic tasks with limited training data.
• An intermediate regularization coefficient value was found to be effective for generalization.
• Comparative simulations showed the proposed model outperformed conventional forward models in goal-directed planning.

Conclusions:

• The PC and AIF frameworks, utilizing a generative model, offer superior generalization for robotic agents compared to traditional forward models.
• Learning a prior distribution within a low-dimensional latent space effectively confines motor plan searches to habituated trajectories, improving goal achievement.