Crucially, reading this blog will reduce your Free Energy.
Animals are able to explore, seek food and remember how to reach it without much effort, yet similar performance with such faculty to plan and imagine shortest paths as well as adapt to changed environment are a long standing challenge in robotics. Robots are now able to reach goal and explore, but they usually require much hand-craft and engineering, as well as a specific adaptation to an…
Intro
Among the human traits that correlate with intelligence, curiosity, as the desire of gaining greater knowledge about the world or ourselves, is one of the most fascinating. With the aid of artificial curiosity, current techniques for control could be equipped with more natural skills, such as self-motivated exploration.
Active Inference is a brain-inspired theory according to which all living beings (aka agents) behave according to a unicum imperative: minimizing their free energy. This means that the brain learns a model of the world, to predict future events, and uses this model to decide what actions to take. For a deeper introduction to Active Inference, you can have a look at our previous post.
People are capable of recognizing and interacting with a plethora of different objects entering in their daily lives. However, object detection and manipulation still proves difficult for current robotic systems. Deep learning methods achieve astonishing performance on detecting known objects, for example using the YOLO object detection algorithm, however these techniques fail when unknown objects…
Simultaneous localisation and mapping (SLAM) has been subject of research for roboticists for decades. Learning a robot to navigate in an uncontrolled, unknown environment is far from trivial. On the one hand, the robot must know where it is on a map, i.e. the “localisation”, on the other hand, the robot should build the map for places it hasn’t visited before, i.e. the “mapping”. Typical SLAM…
Moving from place to place comes without thinking for humans, yet it is something robots continue to struggle with for decades. While there have been tremendous advances in the field of SLAM (simultaneous localisation and mapping, the field that concerns itself with robotic navigation) current approaches still fail to produce stable maps that can be utilised for a long time. We propose a first…
In the case that a robots needs to manipulate, find, or do anything useful, it first needs to localize its target. Living organisms do this by first exploring the area in a seemingly intelligent manner, choosing locations they have not yet visited. When we are searching for something and find a cue that could potentially lead to the solution, we, like any other organism, will investigate this…
When we need to take a decision in life, we typically iterate over our available options, and consider all their pros and cons. For each option, we imagine the possible consequences by mentally rolling out “what-if” scenarios, and evaluate how good the different outcomes match our preferences. Can we have robots operate the same way, by equipping them with a camera sensor?
I often get put in the position in which I have to explain to other researchers, colleagues, friends or family what I am working on, which typically ends up in a primer on active inference. It occurred to me that, besides the scientific literature, there aren’t many sources that explain the subject in layman’s terms. Therefore, I decided to devote this first blog post to the topic.