Autonomous robots need robust perception and navigation in environments where LiDAR, RADAR and RGB-D cameras degrade or fail. MOUSE is an adaptive multimodal fusion framework that learns what each sensor is good for and selects the configuration that suits the conditions at hand, learning several downstream perception tasks jointly rather than one at a time.

See the Projects page for the rest of the group’s current work.

The work runs alongside GPify: leveraging the combined strength of normalizing flow and softmax for an out-of-distribution aware confidence score, written with Rudolph Triebel and Volker Krueger and published open access in the International Journal of Computer Vision (volume 134, article 185). The paper combines normalizing flows with the softmax output of a classifier to produce a confidence score that stays meaningful when the input falls outside the distribution the network was trained on: the situation that makes perception brittle on a robot operating in the world rather than on a benchmark.