A new planning framework lets robot teams guess at obstacles they haven't actually seen yet, and the guesses pay off.
Researchers describe MAGIC, short for Multi-Agent Gaussian belief Inference for Coordination, a framework for Multi-Agent Path Finding: plotting collision-free routes for fleets of robots sharing a space, even when the map is wrong. Classical approaches assume the map is fully known in advance, which falls apart the moment a box falls or a spill appears mid-shift. MAGIC instead treats traversability as uncertain and updates a shared belief about it online, using a Gaussian Markov Random Field and Gaussian Belief Propagation to spread one robot's observation to nearby unobserved cells and feed "detour-aware" costs into standard planners. Across benchmark tests with teams as large as 800 agents, the approach cut the actual distance traveled compared to existing methods in 96.3% of instances.
The trick is spatial inference: prior uncertainty-handling methods, contingency plans or reactive replanning, only act once a robot directly bumps into or sees an obstacle, so each robot learns the map one square at a time. MAGIC's bet is that real obstacles cluster, a puddle, a pallet of fallen stock, a blocked aisle, so one sighting can flag a whole neighborhood of unseen cells before robots drive into it. That matters most for warehouse and logistics fleets, where hundreds of robots share tight corridors and a single spill could otherwise cause a cascade of individually-discovered reroutes.
It's still a simulation result: the paper reports cost reductions, not the compute overhead of running belief propagation live across 800 agents in a real, messier warehouse. Promising math, pending a forklift.