Semantic World Models for Multi-robot systems
Speaker
About this event
Multi-robot systems are increasingly deployed in semi-structured environments such as warehouses and dairy farms, where robots must share space, and coordinate their motion in a robust, safe and scalable way. Model Predictive Control (MPC) provides a flexible framework for generating safe trajectories, but it scales poorly with the number of agents, struggles to represent discrete coordination decisions such as yielding, and depends on reliable and low latency inter-agent communication. This thesis proposes that the decision-making of mobile-robot teams in semi-structured environments can be made more robust and computationally tractable by introducing a shared world model that encodes both the spatial structure and geometry of the environment. By automatically configuring the constraints, objectives, and agent groupings of individual MPC controllers from this shared representation, the need for coupled global planning is reduced while safe and deadlock-free operation is maintained. Four contributions are presented in this thesis. Chapter 2 introduces a property-graph world model built from three map primitives, Areas, Boundaries and Interfaces, and a systematic mapping from these primitives to MPC constraints and objectives, allowing controllers to be reconfigured at runtime and agents to be dynamically grouped only when their semantic area horizons overlap. Chapter 3 adds a discrete claim policy on top of the continuous MPC layer that assigns access rights to areas, ensures deadlock-free operation, and enables task-dependent area occupancy constraints; the approach is validated in simulation and on physical robot hardware. Chapter 4 extends the framework to a distributed ADMM-MPC setting in which the world model selectively activates inter-agent coupling and yielding decisions. Chapter 5 closes the loop by deriving the required world models from Building Information Models using an intermediate representation and a sampling-based roadmap to recover navigable zones. Together, these chapters show that a shared semantic world model is able to configure MPC controllers at runtime, to decouple agent problems through discrete coordination, and to reduce computational and communication requirements compared to baseline methods, leading to multi-robot coordination that is more robust in execution, and configurable at runtime
Host
