This thesis develops a predictive coding approach for haptic object recognition with robotic systems. Unlike conventional bottom-up approaches, it relies on comparing sensory predictions against actual feedback. This enables the robot to integrate diverse sensory information and dynamically adapt its actions. By continuously updating its internal models, the robot can effectively perceive objects through haptic interactions and explore unknown environments stably.

To interact with unknown materials, a haptic exploration framework was developed to estimate major mechanical properties: viscoelasticity, friction coefficient, and restitution coefficient. The identification employs a dual Kalman filter enabling stable interactions and object recognition. Additionally, a dynamic memory was developed to handle novel objects by creating new clusters based on known object knowledge, leading to better clustering results than methods lacking this knowledge-based approach.

To capture nonlinear haptic dynamics during object exploration, a novel bio-inspired recursive neural network approach with the linear summation model (LSM) was developed. The results demonstrated the ability to accurately represent the dynamic interaction, surpassing statistical and linear mechanical representations, as well as reservoir computing approaches, due to LSM’s nonlinear preprocessing and network topology.

An active exploration strategy was then developed to minimise confusion during object classification, initially selecting actions for a broader object understanding and subsequently choosing actions to distinguish likely candidates. This strategy improved classification results and reduced interactions compared to passive or random strategies, surpassing single-metric algorithms in achieving comparable or superior results.

In short, this thesis demonstrated that the presented predictive coding approach can improve robot haptic perception in unknown environments. This involves capturing object characteristics through mechanical and bio-inspired neural representations, anticipating incoming novel objects by forming their clusters, and selecting actions to minimise confusion during classification. The developed approach has wide applicability in surgical, agricultural, and industrial robotic systems, ultimately enhancing interaction with their environment.