From Instruction to Inheritance: Scaling Robot Learning Through Knowledge Circulation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Robot learning, particularly through vision–language–action (VLA) models, has rapidly improved the ability of robots to execute instructed tasks. However, the current mainstream paradigm of robot learning largely depends on one‐way instruction from humans to robots, lacking a cyclical structure in which knowledge acquired by robots is propagated to others. As a result, data collection remains heavily dependent on humans, while learned skills are not effectively inherited, reused, or continuously improved. In this paper, we propose a transition from instruction‐centered robot learning toward an inheritance‐centered framework, in which knowledge and skills circulate from humans to robots, from robots to robots, and even from robots back to humans. To realize such inheritance, it is essential to establish mechanisms that align multiple forms of teaching between robots and other agents, ranging from learning by physical guidance, to learning by watching, and finally to learning by language instruction. From the hardware perspective, this requires appropriate interfaces between humans and robots, while from the software perspective, it requires architectures that connect modalities shared across agents. Furthermore, we propose adaptability, generalizability, and transferability as stepwise milestones toward inheritance and discuss the research challenges and evaluation principles required to achieve scalable knowledge circulation in robot learning.