MAGR-BB slashes multi-agent goal recognition compute by orders

Multi-agent goal recognition is a combinatorial challenge: an observer must simultaneously infer which agents are cooperating and what each team aims to achieve, with the hypothesis space exploding as the number of possible team partitions and goals grows. Real-world applications like drone surveillance and collaborative robotics only expose agent trajectories, forcing the observer to rank team-goal hypotheses purely from behavior. To tackle this, Thiago Thomas, Gabriel de Oliveira Ramos, and Felipe Meneguzzi propose MAGR-BB (Multi-Agent Goal Recognition with Branch-and-Bound). Their method combines a shared team- and goal-conditioned reinforcement learning policy—trained to score how likely a given team-goal hypothesis explains observed behavior—with a factorized branch-and-bound search that prunes unlikely hypotheses without exhaustive enumeration.

On a controlled multi-agent Blocksworld benchmark, MAGR-BB returns the exact same top-ranked hypothesis as exhaustive search throughout the entire trajectory. At the same time, it cuts hypothesis materialization (the number of hypotheses that must be explicitly considered) by orders of magnitude and reduces cumulative recognition runtime substantially. The paper (12 pages, 1 figure, 2 tables) demonstrates that this search-based approach makes real-time multi-agent goal recognition feasible for tasks such as cooperative drone swarms and robotic teams, where inferring intent quickly from movement alone is critical for safety and coordination.