GLOW from signup to working: how to break a one-line task into actions

A friend recommended GLOW to me last week, saying it's the new embodied large model from Knowin AI, and told me to give it a try. I do AI and semiconductor research, and I've been following the embodied line on and off for two months. Two weeks ago I even wrote a post about an active-interaction embodied brain, and my conclusion then was that the solution was usable for pre-research but still a few steps away from product-level delivery. So this time I didn't have high expectations—I just treated it as catching up on a lesson.

From registering to running the first action sequence, it took me less than forty minutes in total.

First, let me be clear about what GLOW is. The full name is Generative Learning of World. You don't need to remember that term. Just understand it as: it lets the robot rehearse the action in its head before actually moving. It has five modules underneath, and I'll go through them in plain language. KnowinDream handles scene generation, like creating a kitchen and the bottles and jars inside it; KnowinWorld handles simulation—if you tell it to pick up a cup, it first calculates what the hand will bump into as it reaches over; KnowinBrain is the brain, responsible for understanding and planning; KnowinAct is the hand, turning the plan into concrete actions; KnowinAgent is the scheduler, stringing the above together.

The official description is "generate the world → understand the world → act on the world → learn from real feedback," forming a closed loop.

What I want to teach is the most practical part: give it a one-sentence task and have it output an action sequence the robot can execute.

There are two paths for the solution, and I read through both in the docs. Path A directly calls the cloud API, installs nothing locally, and you can try it right in the browser—good for validating ideas first. Path B is mounting an edge computing board next to the robot, connected via USB or Ethernet, with the model running locally—good for scenarios where you already have the body and need to cut latency. I went with A this time; for B I only looked at the integration docs.

Here's the specific process.

First, register a developer account and enter the console. Email plus verification code, two minutes. Once inside, first create an application and get a set of API credentials—a long string of characters. Save it in Notepad first; once this page refreshes you won't see it again. Then in the model list, select KnowinBrain. This one handles task planning: you give it a task and a scene, and it gives you steps. Don't pick KnowinAct right away—that one requires connecting to the execution layer, and beginners easily get stuck.

Next is writing the task description, and this step is the watershed. The first time I wrote "tidy up the table," the result was unusable—it didn't know what was on the table, nor what "tidied up" meant as a completion condition. I changed it to "put the white mug on the dining table into the second cabinet on the right; the mug is empty," and the output was normal. The task description must include referable object names and spatial relationships—the more specific the better.

Add a scene description, telling it what's on site and roughly where the objects are. I wrote a paragraph of text by hand; I didn't connect a camera. Then click run and wait for the result. Mine came back in about ten-plus seconds. The output is a numbered sequence of steps, each containing an action type and an object, like move to a certain position, grasp a certain object, place it somewhere. What it gives is an action sequence, not executable code—the conversion layer in between you have to write yourself.

Three pitfalls to mention.

The first is writing the task too vaguely. The fix is to write out the verb, object, position, and end condition completely.

The second is object names not matching up. You write "mug," but the scene description says "cup," and sometimes it treats them as two different things. Just unify the naming.

The third is coordinates. A handwritten scene description can't give precise coordinates, so the output positions are relative, like "right side of the cabinet." To actually land it on the execution layer, you need to hook up a coordinate mapping yourself—I haven't done that part yet.

One word on results. Knowin AI released a technical report at the end of September: KnowinBrain-1.5, on an evaluation platform called Embodied Arena, scored a composite 65.62 on the 2D-Embodied QA leaderboard, ranking first among twenty participating models.

That platform launched in late September, covering dimensions like perception, reasoning, navigation, and planning, and has integrated dozens of models. The score only shows it holds up at the understanding level; whether it can land on a real machine is another matter.

After learning this, you can try two things next. One is swapping the scene description for real-time camera input to see how stable it is handling dynamic frames. The other is digging into the KnowinAct docs and trying to feed the action sequence into a simulation environment to run once—even if it's just a grasping action. Once you get past that step, you've only just touched the edge of the embodied line.

GLOW can help you think clearly about what to do, but how to do it, you still have to hook up yourself.