Multi-agent Trajectory Prediction and Generation with Topological Invariants Enforced by Hamiltonian Dynamics.

We present a planning framework for decentralized navigation in dynamic multi-agent environments where no explicit communication takes place among agents. Our framework is based on a novel technique for computationally efficient multi-agent trajectory generation from symbolic topological specifications. At planning time, this technique allows an agent to generate a diverse set of potential future scene evolutions in the form of Cartesian, multi-agent trajectory representations. The planning agent selects and executes the next action assigned to it from the evolution of minimum cost. The proposed strategy enables an agent to maintain a smooth and consistent navigation behavior that is robust to behaviors generated by heterogeneous agents and to unexpected events such as agents with changing intentions. Simulation results demonstrate the efficacy of our approach in responding to such situations and its potential for real-world applications in crowded environments.