From Finite Embodied Capabilities to Autonomous Knowledge and Method Evolution

Autonomous intelligent systems should be able to acquire knowledge and improve their own learning capabilities through continuous interaction with the environment. However, existing approaches often rely on predefined tasks, fixed learning methods, or externally provided training data, limiting their ability to grow autonomously from minimal initial capabilities. In this paper, we propose a finite-capability-driven autonomous learning framework that actively utilizes sensing, movement, interaction, learning, and thinking to acquire new experiences. The framework transforms experiences into reusable knowledge, applies acquired knowledge to improve task execution, and reinvests saved resources into further learning. It also enables experience-driven evolution of learning and thinking methods through candidate generation, evaluation, and selective retention. Persistent activity scheduling integrates these mechanisms into a continuing autonomous learning process. Results from complementary simulation experiments show that active embodied exploration discovers 25.5 versus 9.7 objects without movement, improves recognition Macro-F1 from 0.588 to 0.634 through active perception, increases cross-object transfer accuracy from 78.82% to 86.54% through rule reuse, and reduces task execution cost by 21.09% compared with a fixed method.