Beyond saccadic movements for interactive perception in Robotics

Resumen

Billions of years of evolution have generated highly complex systems that allow them to adapt to their environments and generate a series of behaviours. Starting from this premise, it seems reasonable to assume that for building adaptive and somewhat intelligent robotic systems, it should be essential to observe biological systems exhibiting such properties. Then, the bio-inspiration concept could be a helpful tool. This thesis is based on the study of the biological model for the generation of saccadic eye movements in humans, which is one of the fundamental pillars for exploring the world around us, and an essential part of the concept of active vision. The adaptation process that guides the learning of this model in a robotic system is based on interactive perception. The morphological characterisation of a robotic system for the execution of this type of movement, along with the development of several internal model proposals allowing the replication of this biological behaviour in a robotic system are some of the main axes of this work. The morphology selected is based on the Helmholtz setup model. This configuration is a binocular system that is a simplified replica of the primate visuo-oculomotor system. Therefore, we perform a preliminary analysis regarding the necessity of encoding the stimulus using binocular or monocular vision. Our result suggest that using the cues from both cameras notably increases the precision of the saccadic movements. The internal model enabling the generation of behaviour from the information coming from the environment through the sensors is described by two different architectures: Feedback error learning and Recurrent architecture. These architectures are based on the principles of interactive perception. Several experiments are conducted to evaluate the characteristics of both architectures, concluding that although the Recurrent architecture offers better performance concerning the accuracy, the time needed to adapt it is a limitation that the feedback error learning architecture does not present. In either case, they are both more precise than suggested in some literature systems and have the quality of being adaptive. In this way, the evolution of the system through interactive perception turns the environment into a fundamental actor in the process, ultimately defining the behaviour of the robot. The estimation of the internal model associated with a given morphology emerges from the interaction with the environment. The adaptive supervised learning techniques presented in this work have an important limitation once the internal model has been learned: any change in morphology or environment involves a learning process. In this thesis, we propose a model to obtain the correlation between morphology and internal model parameters so that a new internal model can be predicted when morphological parameters are modified. Furthermore, we suggest different neural network architectures to address this dimensionally severe regression problem. Using the studied robotic system for generating saccade eye movements, we evaluate the performance of each approach. The best results are achieved for an architecture with parallel neural networks. Our results can be instrumental in state-of-the-art trends such as self-reconfigurable robots, reproducible research, cyber-physical robotic systems, or cloud robotics. Furthermore, internal models would be available as shared knowledge so that robots with different morphologies can readily exhibit a particular behaviour in a given environment. In nature, the transmission of information between generations through genes solves the problem of starting from scratch. Thus, when a living being is born, it does not need to generate its entire internal model to display its behaviours. Instead, it starts from a model encoded by its genes, subsequently modified and adapted by the various environments in which the living being is developing. Based on this premise, the procedure and the machine learning tools used to predict the internal model from morphology allow us to establish the concept of artificial genotype, proposing what an individual and a species are from the point of view of robotics. Finally, we suggest this model exhibits a behaviour that is similar to that of living organisms regarding the concept of reaction norm. To conclude, an algorithm based on microsaccades and head movements during visual fixation is presented. It combines the images generated by these micro-movements with the ego-motion signal in order to compute the depth map. Depth estimation is a challenge for robots and living organisms in their adaptation to evolving environment. We propose a model that abstracts head microdisplacements and microsaccadic movements. A depth map of the initial image can be obtained using the stream of images produced in the visual fixation process as a disturbance in the initial image of the fixation point. The algorithm is tested in a robot eye-in-hand simulation, and, in light of the results obtained, it can satisfactorily estimate the depth map. Also, they corroborate the fact that microsaccades are instrumental in stabilising this estimation. In order to implement this algorithm in a real robotic system, we designed and built a visuo-oculomotor system with a Helmholtz distribution and mounted it on a rotating Sterwart platform enabling us to perform the neck functions. By replicating the fixation movements, we obtained the depth image in both cameras based on the algorithm. In summary, throughout this thesis, we present various algorithms and methodologies with a solid biological inspiration improving the perception and adaptation capabilities of robotic systems in general, being the environment the force that allows the systems to improve their performance.