GLOW (Generative Learning of World)

Overview

GLOW (Generative Learning of World) is a generative learning framework for general-purpose embodied intelligence developed by Knowin. It is not a physical robot but an AI software stack / foundation model that unifies scene understanding, physical reasoning, task planning, and action generation into a single multimodal autoregressive model (KnowinGLOW). The system learns from real and synthetic physical interaction experience, uses world rollouts (KnowinWorld) to broaden exposure to action outcomes, and employs closed-loop execution to turn model decisions into physical robot actions. GLOW supports one-shot teaching via in-context learning, enabling robots to adapt to new tasks from a single human demonstration without updating model parameters. It achieves 62.2% average success across 18 RoboDojo tasks and 86.7% on LIBERO-Pro, and KnowinBrain-1.5 ranks No.1 on Embodied Arena's 2D embodied QA leaderboard.