Event

LSM converts language descriptions into calibrated spatial distributions and integrates them into VL-Map mapping

1 reports · 1 sources

Aryan Naveen and other authors propose and train a Language Sensor Model (LSM), which maps each natural language description together with its scene graph context into a multimodal distribution: mixture weights represent referential ambiguity, and component covariances represent spatial uncertainty. The authors state that, on the VLA-3D benchmark and on a real mobile robot, LSM is the only language predictor whose covariance estimates remain within the calibrated range; after integrating it into the VL-Map probabilistic mapping framework, predictions of target object locations are more accurate, and the probability mass obtained at the true target is about 70% greater than that of the strongest foundation model baseline.