This article presents a dynamic knowledge graph approach for distributed self-driving laboratories (SDLs) within the World Avatar project, aiming to create an all-encompassing digital twin of scientific research laboratories. The approach integrates data and material flows in design-make-test-analyse (DMTA) cycles using ontologies and autonomous agents as executable knowledge components. Data provenance is recorded to ensure findability, accessibility, interoperability, and reusability. The framework is demonstrated by linking two robots in Cambridge and Singapore for real-time collaborative closed-loop optimisation of a pharmaceutically-relevant aldol condensation reaction. The knowledge graph autonomously evolves toward the scientist's research goals, with the two robots effectively generating a Pareto front for cost-yield optimisation in three days. The concept of laboratory automation, reinterpreted as self-driving laboratories (SDLs), has been in existence since the 1960s. SDLs have gained widespread adoption in chemistry, materials science, biotechnology, and robotics, resulting in accelerated scientific discovery and societal development. However, implementing SDLs can be challenging and typically requires a highly specialised team of researchers with expertise in chemistry, engineering, and computer science. Consequently, studies are often conducted by large research groups within a single organisation. In response to global challenges, there is a growing consensus that a paradigm shift towards a globally collaborative research network is necessary. This shift requires decentralising SDLs to integrate different research groups to contribute their expertise towards solving emerging problems. Such decentralisation holds great potential in supporting various tasks ranging from automating the characterisation of epistemic uncertainty in experimental research to advancing human exploration in deep space. The three major challenges in achieving this vision are efficiently orchestrating heterogeneous resources, sharing data across organisations, and recording data provenance following FAIR principles. The proposed architecture enables scientists to set research goals and resource restrictions for a particular chemical reaction and have them trigger a closed-loop process in cyberspace. The process is initiated by the monitoring component, which parses the research goals and requests the iterations needed to achieve the objectives. The iterating component collects prior information about the design space and passes it on to the component that designs the next experiment. The dynamic knowledge graph approach abstracts software components as agents that receive inputs and produce outputs. The flow of data between these components is represented as messages exchanged among these agents. Physical entities can be virtualised as digital twins in cyberspace, enabling real-time control and eliminating geospatial boundaries when multiple labs are involved. This reformulation of the closed-loop optimisation problem as information travelling through the knowledge graph and reflecting their changes in the real world offers a powerful framework for achieving true distributed SDLs. The work demonstrates a proof-of-concept for a distributed network of SDLs enabled by a dynamic knowledge graph. This signifies the first step towards digital research scientists collaborating autonomously. To illustrate the effectiveness of this approach, two robots in Cambridge and Singapore are used to collaborate on a multi-objectiveThis article presents a dynamic knowledge graph approach for distributed self-driving laboratories (SDLs) within the World Avatar project, aiming to create an all-encompassing digital twin of scientific research laboratories. The approach integrates data and material flows in design-make-test-analyse (DMTA) cycles using ontologies and autonomous agents as executable knowledge components. Data provenance is recorded to ensure findability, accessibility, interoperability, and reusability. The framework is demonstrated by linking two robots in Cambridge and Singapore for real-time collaborative closed-loop optimisation of a pharmaceutically-relevant aldol condensation reaction. The knowledge graph autonomously evolves toward the scientist's research goals, with the two robots effectively generating a Pareto front for cost-yield optimisation in three days. The concept of laboratory automation, reinterpreted as self-driving laboratories (SDLs), has been in existence since the 1960s. SDLs have gained widespread adoption in chemistry, materials science, biotechnology, and robotics, resulting in accelerated scientific discovery and societal development. However, implementing SDLs can be challenging and typically requires a highly specialised team of researchers with expertise in chemistry, engineering, and computer science. Consequently, studies are often conducted by large research groups within a single organisation. In response to global challenges, there is a growing consensus that a paradigm shift towards a globally collaborative research network is necessary. This shift requires decentralising SDLs to integrate different research groups to contribute their expertise towards solving emerging problems. Such decentralisation holds great potential in supporting various tasks ranging from automating the characterisation of epistemic uncertainty in experimental research to advancing human exploration in deep space. The three major challenges in achieving this vision are efficiently orchestrating heterogeneous resources, sharing data across organisations, and recording data provenance following FAIR principles. The proposed architecture enables scientists to set research goals and resource restrictions for a particular chemical reaction and have them trigger a closed-loop process in cyberspace. The process is initiated by the monitoring component, which parses the research goals and requests the iterations needed to achieve the objectives. The iterating component collects prior information about the design space and passes it on to the component that designs the next experiment. The dynamic knowledge graph approach abstracts software components as agents that receive inputs and produce outputs. The flow of data between these components is represented as messages exchanged among these agents. Physical entities can be virtualised as digital twins in cyberspace, enabling real-time control and eliminating geospatial boundaries when multiple labs are involved. This reformulation of the closed-loop optimisation problem as information travelling through the knowledge graph and reflecting their changes in the real world offers a powerful framework for achieving true distributed SDLs. The work demonstrates a proof-of-concept for a distributed network of SDLs enabled by a dynamic knowledge graph. This signifies the first step towards digital research scientists collaborating autonomously. To illustrate the effectiveness of this approach, two robots in Cambridge and Singapore are used to collaborate on a multi-objective