Knowin Launches GLOW Architecture to Enable Robots to Learn Tasks in One Demonstration
For a long time, teaching a robot a new task has required collecting fresh data, writing fixed procedures, or running targeted training. Such methods suit highly deterministic tasks but struggle with the changing objects, environments, and needs of real-world settings.
Recently, the release of a new-generation general model, GPT-6 Astra, sparked broad discussion and rethinking of embodied intelligence technology routes. Some observers argue that general models are demonstrating the ability to control robots, potentially disrupting existing embodied AI approaches.
At the same time, the industry is converging on a consensus: autoregressive architectures are becoming the convergent model design, and the core competition in embodied intelligence will shift toward data and learning methods.
On September 24, Knowin released a technical report on GLOW, a generative learning architecture for general embodied intelligence. The report systematically introduces GLOW’s upgraded architecture and presents a viable path for turning general cognitive capabilities into reliable physical action. Its core breakthrough is enabling robots to learn a task from a single demonstration.
GLOW aims to give robots general embodied abilities reusable across objects, environments, and tasks, rather than learning a separate policy for each task. Users need not break down requirements into machine commands or be limited by a factory-written function list. A robot can understand a task through one complete operation and translate visuals, object relationships, state changes, and execution feedback into action. This capability has been validated in household tasks ranging from unpacking and storage to watering, wiping tables, and bartending.
The “learn from one demonstration” ability stems from GLOW’s upgraded technical path. GLOW unifies its original Brain and Act capabilities into a multimodal autoregressive model, KnowinGLOW, enabling visual understanding, spatial reasoning, task planning, and action generation to work together in one model. KnowinDream generates physical interaction experience, KnowinWorld simulates possible outcomes of actions, and KnowinAgent organizes execution around task memory, tool calls, feedback, and replanning. This forms a complete chain from experience learning and task understanding to action feedback.
Evaluation results and the route report disclose GLOW’s scores on RoboDojo, LIBERO-Pro, and Embodied Arena. Experiments show that GLOW can not only understand physical scenes but also translate that understanding into effective task execution. Under the report’s evaluation settings, GLOW achieved leading results in multiple mainstream international embodied benchmarks, demonstrating capabilities in spatial understanding, physical interaction, and task execution. Knowin continues to advance this route: training models with scaled physical experience, coordinating cognition and action through a unified architecture, and adapting to environmental change through closed-loop execution. General models strengthen knowledge, understanding, and reasoning, while specialized embodied models center on physical interaction to build a complete capability from scene understanding to autonomous action.