Inhabitant: training-free LLM agent completes 74/75 long-horizon household tasks

Yao Zhang · hf · 2026-11-20

Aalto University's Inhabitant framework lets an LLM agent plan and execute long-horizon human-scene interactions (grasp, kick, throw) by specifying whole-body motions spatially, measuring deviations in the scene, and revising control — with no training. On the accompanying Inhabit benchmark the authors report 74/75 long-horizon household tasks completed. By Yao Zhang, Zhuchenyang Liu and Yu Xiao.

Original post →

More from Embodied

Embodied channel →