Knowin, a company specializing in embodied AI, published a technical report on September 24th about GLOW, which stands for Generative Learning of World. This framework is designed to allow domestic robots to learn a new task based on just one complete human demonstration. The company claims that the acquired skill can then be applied to various objects, environments, and tasks without the need for retraining or changing model parameters.
GLOW consists of four main components. KnowinGLOW is a unified multimodal autoregressive model that integrates visual understanding, spatial reasoning, task planning, and action generation, directly outputting commands for grasping and end-effector pose. KnowinDream creates physically grounded experiences in domestic settings, transforming daily routines into 3D scene simulations, where lighting, materials, objects, and camera angles are varied. KnowinWorld evaluates the probable outcomes of proposed actions, ranking them by task progress, physical consistency, achievability, and risk. The KnowinAgent, also known as Harness, acts as the execution environment, managing memory, tool calls, spatial queries, execution feedback, and replanning during long-duration tasks.
The one-shot learning method relies on in-context learning. A demonstration encoder compresses a video of a single demonstration into a reusable skill context, which the model combines with current camera views, language instructions, robot state, and task history, while the model weights remain unchanged. Knowin demonstrated the application of this technology in scenarios such as storing boxes, watering plants, wiping tables, and mixing drinks, with robots adapting to changes in object placement or orientation. Execution feedback plays a significant role: when tasked with an item in one drawer, the robot requested a movement of 24.3 mm but only achieved 1.6 mm, detected an obstacle, and changed its approach.
Knowin presents several testing results on its own benchmark setups. On 18 simulated RoboDojo tasks, GLOW showed an average success rate of 62.2% and a score of 71.1, significantly higher than the baseline GPT-6 model (22.2% and 29.8, respectively). Across six LIBERO-Pro perturbation conditions, it achieved an average of 86.7% as a single policy, which is the highest among single-policy models in the comparative table, although some multi-policy agent systems showed better results. The KnowinBrain-1.5 model ranks first among 20 models in the Embodied Arena response rating for embodied questions, with an overall score of 65.62 across nine benchmarks.
The report also notes limitations, including instances of success checks triggering even when the object was still being grasped, and that step counting does not allow for strict speed comparisons. Knowin was founded in August 2025 and focuses on developing generalized embodied models for consumer home use. Its research team comprises over 200 people, and the company's first product, Knowin-X1, is undergoing validation before mass production.