In this paper we investigate the active inference framework as a means to enable intelligent behaviour in artificial agents. Active might underpin way organisms act and observe real world. inference, agents order minimize their so called free energy, or prediction error. Besides being biologically plausible, has been shown solve hard exploration problems various simulations. However, these simulations typically require handcrafting generative model for agent. Therefore propose use recent...

Simultaneous localization and mapping (SLAM) represents a fundamental problem for autonomous embodied systems, which the hippocampal/entorhinal system (H/E-S) has been optimized over course of evolution. We have developed biologically-inspired SLAM architecture based on latent variable generative modeling within Free Energy Principle Active Inference (FEP-AI) framework, affords flexible navigation planning in mobile robots. primarily focused attempting to reverse engineer H/E-S "design"...

Learning to take actions based on observations is a core requirement for artificial agents be able successful and robust at their task. Reinforcement (RL) well-known technique learning such policies. However, current RL algorithms often have deal with reward shaping, difficulties generalizing other environments are most sample inefficient. In this paper, we explore active inference the free energy principle, normative theory from neuroscience that explains how self-organizing biological...

Active inference is a process theory of the brain that states all living organisms infer actions in order to minimize their (expected) free energy. However, current experiments are limited predefined, often discrete, state spaces. In this paper we use recent advances deep learning learn space and approximate necessary probability distributions engage active inference.

Occlusions, restricted field of view and limited resolution all constrain a robot's ability to sense its environment from single observation. In these cases, the robot first needs actively query multiple observations accumulate information before it can complete task. this paper, we cast problem active vision as inference, which states that an intelligent agent maintains generative model acts in order minimize surprise, or expected free energy according model. We apply object-reaching task...

Simultaneous localization and mapping (SLAM) represents a fundamental problem for autonomous embodied systems, which the hippocampal/entorhinal system (H/E-S) has been optimized over course of evolution. We have developed biologically-inspired SLAM architecture based on latent variable generative modeling within Free Energy Principle Active Inference (FEP-AI) framework, affords flexible navigation planning in mobile robots. primarily focused attempting to reverse engineer H/E-S ‘design’...

The human intrinsic desire to pursue knowledge, also known as curiosity, is considered essential in the process of skill acquisition. With aid artificial we could equip current techniques for control, such Reinforcement Learning, with more natural exploration capabilities. A promising approach this respect has consisted using Bayesian surprise on model parameters, i.e. a metric difference between prior and posterior beliefs, favour exploration. In contribution, propose apply latent space...

Biologically inspired algorithms for simultaneous localization and mapping (SLAM) such as RatSLAM have been shown to yield effective robust robot navigation in both indoor outdoor environments. One drawback however is the sensitivity perceptual aliasing due template matching of low-dimensional sensory templates. In this paper, we propose an unsupervised representation learning method that yields latent state descriptors can be used RatSLAM. Our sensor agnostic applied any modality,...

This paper investigates unsupervised learning of low-dimensional representations from FMCW radar data, which can be used for multiple downstream tasks in a drone navigation context. To this end, we release first-of-its-kind dataset raw ADC data recorded mounted on flying an indoor environment, together with ground truth detection targets. We show that, by utilizing our learned representations, match the performance conventional processing techniques while training models different input...

As warehouses, storage facilities and factories become more expanded equipped with smart devices, there is a substantial need for rapid, intelligent autonomous detection of unusual potentially hazardous situations, also called anomalies. In particular Autonomous Guided Vehicles (AGVs) that drive around these premises independently, unforeseen obstructions along their path-e.g. cardboard box in the middle corridor or bumps floor-and sudden unexpected actions executed by personnel-e.g. someone...

Frequency-modulated continuous-wave (FMCW) radar is a promising sensor technology for indoor drones as it provides range, angular well Doppler-velocity information about obstacles in the environment. Recently, deep learning approaches have been proposed processing FMCW data, outperforming traditional detection techniques on range-Doppler or range-azimuth maps. However, these come at cost; each novel task neural network architecture has to be trained high-dimensional input stressing both data...

Scene understanding and decomposition is a crucial challenge for intelligent systems, whether it object manipulation, navigation, or any other task. Although current machine deep learning approaches detection classification obtain high accuracy, they typically do not leverage interaction with the world are limited to set of objects seen during training. Humans on hand learn recognize classify different by actively engaging them first encounter. Moreover, recent theories in neuroscience...

Active inference is a theory that underpins the way biological agent's perceive and act in real world. At its core, active based on principle brain an approximate Bayesian engine, building internal generative model to drive agents towards minimal surprise. Although this has shown interesting results with grounding cognitive neuroscience, application remains limited simulations small, predefined sensor state spaces. In paper, we leverage recent advances deep learning build more complex models...

Drone navigation in GPS-denied, indoor environments, is still a challenging problem. As drones can perceive the environment from richer set of viewpoints, simultaneous localization and mapping (SLAM) becomes more complex, while having stringent compute energy constraints. To tackle that problem, this research displays biologically inspired deep-learning algorithm for monocular SLAM on drone platform. We propose an unsupervised representation learning method yields low-dimensional latent...

Historically, artificial intelligence has drawn much inspiration from neuroscience to fuel advances in the field. However, current progress reinforcement learning is largely focused on benchmark problems that fail capture many of aspects are interest today. We illustrate this point by extending a T-maze task for use with algorithms, and show state-of-the-art algorithms not capable solving problem. Finally, we out where insights could help explain some issues encountered.

When collaborating with multiple parties, communicating relevant information is of utmost importance to efficiently completing the tasks at hand. Under active inference, communication can be cast as sharing beliefs between free-energy minimizing agents, where one agent's get transformed into an observation modality for other. However, best approach transforming observations remains open question. In this paper, we demonstrate that naively posterior give rise negative social dynamics echo...

Recently, 3D Gaussian Splatting has emerged as a promising approach for modeling scenes using mixtures of Gaussians. The predominant optimization method these models relies on backpropagating gradients through differentiable rendering pipeline, which struggles with catastrophic forgetting when dealing continuous streams data. To address this limitation, we propose Variational Bayes (VBGS), novel that frames training splat variational inference over model parameters. By leveraging the...