REAL: an embodied agent that explores real rooms and asks clarifying questions, accepted at ECCV 2026

jiqizhixin · x · 2026-09-23

REAL, accepted at ECCV 2026, targets two over-idealized assumptions in embodied simulation: fully observable environment state and perfectly clear user intent. The framework unifies active exploration, visual goal localization, natural language intent disambiguation, and physical manipulation into one decision loop, using first-person RGB as the primary perception input. In a real-room demo, given only "get me something to eat," the robot explores, finds a bread and two donuts, asks which one the user wants, and acts on the feedback — no privileged information, no explicit instructions, sim-to-real deployable.

Original post →

More from Embodied

Embodied channel →