Our society is increasingly in need of smarter and adaptive artificial systems that can support us in different tasks. However, programming robots to work in dynamic human environments is costly and challenging. Out of laboratories, their functioning mostly relies on pre-engineered behaviours and the way they perceive things is strongly biased by their algorithmic design. This limits their adaptivity and efficacy. On the contrary, humans incrementally acquire capabilities, and refine and adapt them over time. Experience and factors such as emotions and moods shape the way we learn and further perceive ourselves and the world around us.
This project studies developmental processes with the aim of producing more adaptive robots. It researches learning mechanisms, exploration behaviours and artificial curiosity to let robots autonomously acquire knowledge about their bodily capabilities. Experience can teach robots, like infants, what to expect from further interactions with the surroundings. This research studies the formation of predictive capabilities, i.e the capability to anticipate upcoming sensory information. Current theories on brain functioning consider this as a fundamental computational component for motor control, cognitive development and adaptivity in humans.
In particular, this project had three objectives:
O1. implementing developmental mechanisms for the acquisition of predictive capabilities in robots. Mechanisms for curiosity-driven exploratory behaviours and continual learning, based on deep neural networks and memory consolidation techniques and in the context of multi-modal and high-dimensional sensory spaces, have been developed. These have allowed different robotic platforms to incrementally form predictive internal models.
O2. Enhancing robot perception through predictive processes. Anticipating sensory observations can allow modulating the focus of attention towards the most relevant part of the input for the task at hand. For instance, expected or non-relevant portions of the visual input can be filtered out, leaving thus the novel or most informative part of the scene in the field of view. Moreover, self-regulatory mechanisms can be enabled by monitoring the prediction performance during learning, e.g. modulating the allocation of computational resources or arbitrating internal goals, depending on the novelty of incoming sensory data.
O3. Studying predictive processes as a prerequisite for an artificial self. We perceive and anticipate the results of our own actions differently than those of others. The mismatch between expectations and observations - i.e. the prediction error - has been argued to be a marker characterising the experience of the “self”. This project has studied the role of prediction errors and their dynamics in robot self-perception. Moreover, it produced interdisciplinary theoretical work on the topic of self-perception in humans and robots.
Ultimately, predictive processes can represent a low-cost strategy to enhance current sensing technologies, and studying them can also help shedding light on phenomena such as (self-)perception in humans and on their possibility in artificial systems.
