Active Inference for Tactile Sensor Control

Abstract

This thesis presents an active inference-based agent for decision making and control of a tactile robotic system. We adapt an active inference framework from a modelling tool, to an active control system through the introduction of decision thresholds and finite horizon iterations. The agent is successfully able to complete localisation tasks to a high degree of accuracy on a live robot, and is able to overcome issues with incorrectly classified data to localise correctly. The agent displays behavioural trends that can be tuned via the input matrices that encode the agent’s understanding of the world, without requiring human bias to tune these matrices. These trends show that the agent will select advantageous movement policies that will move the robot away from areas of confusion towards states with high information content. We also show that the world can be characterised into discrete shape states, which would allow for the agent to act in unknown and unstructured environments without requiring relearning a model of its surroundings. Chapter 4 presents the initial implementation for the offline testing of the active inference control system. These initial results show proof of concept for this type of control system, and introduce adaptations to move the active inference framework from a modelling tool to an active control system. Chapter 5 presents the real time implementation of the active inference control system for controlling a tactile robot in real time. The results show the agent is able to act in an optimal manner seeking out areas of high information in its goal to localise. Chapter 6 presents research towards creating shape-based state estimation for tactile robots, so that the active inference agent can be adapted for use in unstructured environments. These shape states are able to generalise the overall shape of the world, rather than relying upon more specific location-based states. Overall, this thesis presents an active inference-based control system for tactile robots that is able to select optimal control policies using minimal information and without requiring human bias.

Date of Award	12 Jun 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Nathan F Lepora (Supervisor) & Laurence Aitchison (Supervisor)