Multi-robot systems (MRS) are coordinated ensembles of two or more robotic agents that collaborate to accomplish tasks beyond the reach of a single robot, using communication, task allocation, and shared world models. They span architectures from tightly coupled homogeneous fleets to loosely coupled heterogeneous teams where robots with different capabilities complement each other. Key research challenges include task and motion planning, collision avoidance, fault tolerance, and communication bandwidth constraints in dynamic environments. MRS find application in search and rescue, warehouse automation, precision agriculture, and planetary exploration.
Content
- The field emerged in the late 1980s from distributed AI and the recognition that parallel embodied agents could solve problems intractable for single robots. Early work focused on simple biological analogies — ant colonies and bee swarms — before maturing into rigorous task-allocation and market-based coordination frameworks through the 1990s and 2000s.
- Coordination in MRS operates across several layers: perception fusion aggregates sensor streams from multiple agents; task allocation assigns objectives using auction-based or optimisation algorithms; motion planning ensures collision-free trajectories respecting both individual dynamics and inter-robot constraints; and communication protocols propagate state updates within latency budgets.
- Industrial applications drive commercial adoption. Warehouse systems such as those deployed by Amazon and Ocado use hundreds of ground robots coordinated in real time. Search-and-rescue MRS can penetrate disaster environments with heterogeneous aerial-ground teams where no single platform suffices. Agricultural fleets spray, harvest, and monitor fields with a fraction of the human labour previously required.
- In 2024-2025, large language models are being integrated as high-level task planners for heterogeneous robot teams, enabling natural-language mission specification. Foundation models for robotics pre-trained on diverse manipulation and navigation data are reducing per-robot sample complexity. Sim-to-real transfer using photorealistic simulation is accelerating the deployment of multi-robot coordination policies trained entirely in software.