Real-World Robot Control Based on Contrastive Deep Active Inference with Learning from Demonstration

Abstract

Robots are expected to achieve their goals through human-like perception and action. Deep active inference based on the free-energy principle (FEP) is a promising approach. However, most studies have only considered toy problems in simulated environments. To overcome this limitation, we propose another deep active inference framework for real robots. This framework consists of a world model, an action model, and an expected free energy (EFE) model. The world model ensures adequate perception of the environment by minimizing contrastive variational FE, while the action model generates adaptive actions through imitation learning by minimizing contrastive EFE estimated by the EFE model, where each energy is adapted from the original one for contrastive learning. We show that a real robot with the framework can successfully perform a reaching task in both learned and unlearned environments. These results highlight the utility of the FEP with contrastive learning for real-world robot control.