2025 Theses Doctoral

Self-Supervised Learning for Self-Modeling Robots

Hu, Yuhang

Humans form internal representations of their bodies and capabilities through constant interaction with their environment. This ability motivates the creation of autonomous robots that can learn and adapt similarly. The central question driving this dissertation is how to equip robots with comparable, self-contained learning processes that unfold without continuing human supervision.

To address this need for human-like adaptability, we propose a framework consisting of two core methodologies: (1) Self-modeling, where deep neural networks capture a robot’s intrinsic attributes, from morphology, kinematics, and dynamics to human-like cognition. (2) Self-supervised learning, in which the robot’s own sensory streams furnish the training data, thus removing the reliance on manually annotated datasets. By combining these methods, our approach lays the groundwork for robots to evolve their knowledge and skills over time, mirroring the fluid developmental progression observed in humans.

Experimental results demonstrate that robots equipped with self-models and self-supervised learning capabilities exhibit higher resilience, faster adaptation, and greater autonomy, even in the face of morphological changes or hardware malfunctions. These findings conclude that combining self-modeling with self-supervised learning establishes a promising foundation for lifelong robotic learning, paving the way for highly robust, improving autonomous systems capable of operating in unstructured environments.