Knowin Releases GLOW Generative Learning Architecture for General-Purpose Embodied Intelligence

On September 24, Knowin released a technical report on GLOW, a generative learning architecture designed for general-purpose embodied intelligence. The report details an upgraded architecture aimed at giving robots the ability to learn a task from a single demonstration and reuse that capability across different objects, environments, and tasks.

According to the report, GLOW unifies the company's previous Brain and Act capabilities into a multimodal autoregressive model called KnowinGLOW, allowing visual understanding, spatial reasoning, task planning, and action generation to work within the same model. KnowinDream generates physical interaction experience, KnowinWorld simulates likely outcomes of actions, and KnowinAgent organizes execution around task memory, tool use, feedback, and replanning. Together they form a pipeline from experience learning and task understanding to action feedback.

Knowin said users do not need to convert requests into machine instructions or rely on a fixed set of out-of-the-box functions. A robot can understand a task after one complete demonstration and translate visual information, object relationships, state changes, and execution feedback into action. The company said the approach has been validated in household tasks such as unpacking and tidying, watering plants, wiping tables, and mixing drinks.

The report also disclosed GLOW's results on benchmarks including RoboDojo, LIBERO-Pro, and Embodied Arena. In the report's evaluation settings, GLOW achieved leading results on multiple international embodied intelligence benchmarks, demonstrating capabilities in spatial understanding, physical interaction, and task execution.