Swarm intelligence is a research field that studies the behavior of large animal societies such as insects, flocks of birds and schools of fish. These natural systems display common properties such as adaptivity, parallelism, robustness, and scalability. Designing artificial systems that match the performance of their natural counterparts is a challenging but fruitful goal. Swarm robotics studies methods to apply the principles of swarm intelligence to complex coordination problems for large robot groups. Swarm robotics systems are typically characterized by fully decentralized control, local interactions among individuals, redundancy, and absence of predefined roles. Future swarm robotics systems hold the promise to provide solutions for problems such as exploration in hazardous areas, construction, and search-and-rescue. Designing effective swarm robotics systems remains extremely challenging. Most of the work in the field is developed in simulation due to the high cost of running large-scale real experiments. Various techniques exist to model swarm robotics systems, but physics-based simulations are currently the most widespread. Before 2006, experiments in swarm robotics focused mainly on navigation-based activities (exploration, pattern formation, flocking) in 2-dimensional environments. With the Swarmanoid project (2006-2010), research moved towards complex swarms composed of different kinds of robots acting in 3D. The existing physics-based simulators did not provide the necessary features to support this wide spectrum of possible research directions. For this reason, I designed ARGoS - a state-of-the-art, generic, physics-based, multi-robot simulator. The ARGoS architecture is designed to provide high degrees of efficiency and accuracy, while proving sufficiently flexible to suit the needs of any experiment in swarm robotics. ARGoS is designed upon five principles: modularity, composability, the possibility to run multiple physics engines, the possibility to tune the accuracy of every aspect of the simulation, and parallelism. Experiments show that ARGoS can perform accurate physics-based simulations of more than 10,000 robots in real time on a modern desktop computer.