Miniaturisation will be the key challenge for the next decade in numerous industrial fields, such as microelectronics, optics and biomedical engineering. Although most of their products already achieve footprints of some square millimeters, the trend towards the integration of a maximum number of elements in a minimal volume requires even more compact components. This tendency creates a growing need for industrial robots able to perform micromanipulation and microassembly tasks with a submicrometric precision. Nonetheless, the design of such machines is nowadays costly, both in time and money, mostly because of the twofold complexity of their development: first, from a kinematic standpoint, the use of a parallel structure consists in a particularly interesting approach to build ultra-high precision robots. However, the synthesis of such a kinematics proves especially challenging for machines presenting more than 3 degrees of freedom. Moreover, the resulting robots are scarcely flexible: if the industrial specifications are modified, which for example necessitates to add a degree of freedom or to change the position of a rotation centre, the design process has to be restarted, often from the very beginning. The second challenge consists in the mechanical design of flexure-based mechanisms: flexure hinges are joints which are based on the elasticity of the matter. They allow to perform motions which are without friction, backlash and wear; their use is thus mandatory to achieve the aimed submicrometric precision. Albeit the synthesis of planar and low-degree of freedom structures is now widely investigated, the development of a whole tridimensional flexure-based robot is still infrequent, especially in the industrial context. This thesis thus introduces a modular design methodology which significantly reduces the time-to-market of ultra-high precision robots. This procedure can be compared to a robotic Lego, where a finite number of conceptual building bricks allows to easily design and modify parallel robots. Furthermore, this work shows that the machines resulting from this approach present similar or even improved performances compared to robots developed more traditionally. The key aspect of this thesis consists in the concept of modular kinematics, which aims at facilitating the synthesis of parallel kinematics thanks to solution catalogues. At this step of the methodology, the conceptual building bricks and the kinematics are totally independent from any mechanical design: they can thus be used to synthesise a large variety of robots, from machine-tools to microscale robots. An exhaustive conceptual solution catalogue groups all kinematics generated by the combination of the building bricks. Then, a reduced solution catalogue for ultra-high precision is proposed: based on selection criteria linked with the design and machining of flexure-based mechanisms, it allows to reduce the total number of solutions and thus facilitates the practical use of the concept. The second part of this work details the mechanical design of the building bricks, whose main challenge consists in increasing the ratio between the working ranges of the mechanisms and their overall size. One or more flexure-based solutions have been developed for each motorised brick: a special emphasis is given to the original use of a Remote Centre of Motion, which allows to achieve high rotation angles while drastically reducing parasitic translations. The development of a standardised actuation sub-brick, common to all motorised bricks, introduces a new level of modularity, thus increasing even more the flexibility of the methodology. As for non-actuated bricks, original designs and uncommon uses of well-known mechanisms are proposed. A case study on a 5-degree of freedom robot, Legolas 5, finally illustrates the practical use of the methodology: first, the selection in the solution catalogue of a kinematics adapted to the specifications of the robot is detailed. Then, the development of the Legolas 5 prototype highlights the mechanical design of the necessary building bricks, as well as assembly subtleties, such as force alignment and gravity compensation, which allow to shrewdly design a high-performance robot. The measurements of this machine have shown motion resolution and repeatability of 50 nm in translation and 1.9 µrad in rotation (limited by the sensor resolution). This case study has generated the Legolas family, a new family of ultra-high precision parallel robots, which notably includes the orthogonal version of the Delta kinematics: using only 6 of the conceptual building bricks, one solution can be built for each of the 19 possible robot mobilities. The promising characterisation of the Legolas 5 tends to suggest that the robots from this family will be interesting candidates to fulfill the upcoming need for quickly designed and high-performance industrial ultra-high precision machines.