• biologically-inspired robots
• free-energy principle
• active inference
• manipulator robot
• active sensing

The objective of this thesis is twofold. First, a formalization of the Active Sensing problem is made with a focus on its state of the art in the context of robotic manipulators, going on to synthesize existing research into a coherent and unified framework. Second, a focus is made on Active Inference, a branch of Active Sensing. This is presented both in a purely theoretical way and through an implementation for controlling a PUMA 560 robot in a Matlab/Simulink environment.
Robotics is, especially in recent decades, constantly developing. Indeed, it is becoming increasingly evident how society is evolving toward a future in which robots have a marked presence and in a wide variety of fields. Hence, there is a growing need to find new ways for robots to achieve as much autonomous behavior as possible. In this context, Active Sensing takes shape.
Active sensing is a technique used in robotics to actively acquire information from the surrounding environment in order to make decisions autonomously and act accordingly. It requires algorithms that integrate perception, learning and control into a single system and thus are able to take into account the physical and geometric properties of objects, internal structure and interactions with the surrounding world. In addition, they must be able to adapt to changes in the environment and different situations.
Algorithms based on active inference fall within the set of active sensing mainly in the field of active exploration. Active Inference is a general framework for perception and action that is gaining importance in computational and systems neuroscience, but is less well known outside these fields. It is based on Karl Friston's Free-Energy principle: biological systems are thermodynamically open, in that they constantly exchange energy and entropy with the environment, exhibiting self-organizing behavior. However, organisms interact with the environment, changing their position within it or their relationship to it; this is done in a way that allows the organism to maintain its states within limits that do not affect its physical structure.
Free-Energy is thus the mathematical formulation of how adaptive systems manage to resist the natural tendency of disorder (entropy). It is understood as the difference between the brain's expectations of the information it receives from the environment and its internal predictions generated on the basis of prior knowledge. Minimizing this energy means reducing this discrepancy.
Assuming that we have position and velocity sensors, that Gaussian noise on these is uncorrelated, and that the states are the positions of the joints (so the arm is controlled in joint space by Free Energy minimization), all the theory presented in the Active Inference part is converted to define a control scheme suitable for a PUMA 560 robot. Given an initial state, the goal of such control is to bring the robot to a predefined final configuration (understood as joint angles).
The simulation (done with different noises on the sensors), shows that the Active inference control turns out to be extremely easy to model on a manipulator and also very adaptive. Joint convergences occur after extremely short intervals and no oscillations are present. In addition, the number of variables to be tuned in the AIC is constant and does not depend on the number of DoFs of the robot (unlike a classical adaptive control): it’s always 6. Finally, the AIC responded well even to fairly pronounced perturbations.