Particle.news
Download on the App Store

TUM Robot Finds Misplaced Items on Voice Command Using LLM-Guided Search

A two-part system that pairs selective 3D mapping with language-model semantics concentrates scans on likely locations, yielding faster and more efficient results in lab tests.

Overview

  • Developed at TUM’s Learning Systems and Robotics Lab, the mobile platform builds centimeter-accurate room maps by combining camera images with depth data.
  • An active update mode prioritizes rescanning surfaces that often change, such as tables and countertops, while de-emphasizing largely static areas.
  • A large language model provides commonsense relations about objects and interprets voice queries so the robot can target the most probable locations first.
  • In controlled evaluations, the prototype found objects 14% faster and searched about 30% more efficiently, with a reported 95% accuracy for detecting environmental changes.
  • The work by Benjamin Bogenberger et al. is documented in IEEE Robotics and Automation Letters (DOI: 10.1109/LRA.2026.3656790), and the team aims to add manipulators to open drawers and doors for hidden-item retrieval.