Report
A robot-mapping method treats language as a sensor alongside vision
The researchers say their system weighs possible meanings of a location clue before combining it with camera observations.
TLDR
In their CoRL 2026 paper, the researchers describe turning remarks such as “I left my backpack on the table” into probabilities over possible locations, then combining those estimates with vision. They report about 70% more probability mass on the true target than the strongest foundation-model baseline in tests involving the VLA-3D benchmark and a mobile robot.
Combined views
3
2 Sources, first seen ago
