Awesome collaboration with @hocherie.bsky.social, @bradygmoon.bsky.social, A.Parandekar, N. Harutyunyan, C. Wang, K. Sycara, G. Best, S. Scherer. @airlabcmu.bsky.social

We evaluated MapEx on a KTH floorplan dataset simulator. We show MapEx outperforms state-of-the-art map prediction-based exploration baselines. 5/n

MapEx generates multiple predicted maps from the observed maps, from which we compute mean and variance maps.

We take the variances of predicted maps and estimated visible area into consideration to compute the information gain of a given viewpoint for exploration planning. 4/n

Real-world indoor environments often have structure and repeating patterns. To leverage this, we train LaMa, an inpainting network, to predict unknown areas from partial observations—enabling our map predictor to infer beyond what's seen. 3/n

Tl;dr: Predict multiple maps → estimate uncertainty and sensor coverage → explore to build accurate maps faster.

We introduce MapEx, a new exploration framework that uses predicted maps to form probabilistic sensor models as information gain metric for exploration planning. 2/n