Robots exploring unknown indoor spaces face a fundamental problem: they must decide where to go while most of the environment is still unseen. Recent predictive approaches often rely on learned models that need offline training and repeated inference during exploration. The MAPLE method instead builds a training-free, online generative prior using a Mondrian process, conditioned on geometric cues already present in the robot's partial map.
Rather than producing a single predicted occupancy map, MAPLE generates multiple plausible map hypotheses and evaluates frontier gain separately within each. This preserves differences in predicted visibility across hypotheses, which the authors argue is important for making good exploration decisions. The lightweight procedural construction allows predictive exploration even with limited computational resources.
In experiments across KTH floor plans and real-SLAM environments at multiple sensor ranges, MAPLE consistently improved over non-predictive exploration and remained competitive with learned predictive approaches. The authors also report a notable finding: more accurate map prediction did not necessarily lead to better exploration, highlighting that how predictions are represented and used matters as much as their raw accuracy.