Stabilizing the Convergence of Pixel-Based Deep Active Inference Controllers Using Adaptive Smoothing Filters

In recent years, active inference has gained attention in robot control owing to its adaptability to environmental changes. However, its reliance on gradient descent of variational free energy offers no guarantee of convergence to an optimal solution. In this study, we propose an approach that applies a smoothing filter to a pixel-based active inference controller to mitigate the risk of local minima. By smoothing the observed, predicted, and target values, the free energy function becomes smoother, yielding a broader distribution of gradients toward the target, thereby reducing the risk of being trapped in the local minima. In addition, in order to prevent excessive smoothing from eliminating the gradient of the free energy function, we also proposed a method for dynamically adjusting the intensity of smoothing based on prediction and target errors. To evaluate the effectiveness of our method, we applied it to two simulation environments: a simple object-tracking task using a 3-degrees-of-freedom camera, and a robot control task using a 2-degrees-of-freedom robotic arm, and compared it with the conventional active inference controller as a baseline. The experimental results demonstrate that the proposed approach achieves improved convergence performance over the conventional method.