
How I hooked up a motion brain to my own robot body
I compared the VLA path with this world action model setup, and actually ran it.
VLA's approach is to take a language model that already knows how to look at images and talk, then feed it a bunch of "see this scene, do this action" data, transferring language understanding to action output. World action models go the other way — starting from the action itself. This difference matters a lot to robot OEMs. One piece in the materials puts it bluntly: a top-tier robot brain on an H100 takes about 300 milliseconds to make a decision — enough for you to blink twice. On a production line, those 300 milliseconds count toward cycle time.
These past few days I ran the simplest pick-and-place task with the Moushen Intelligence setup, from registration to producing an action sequence that can be replayed in simulation — about two hours total. Below is the order I went through; the interface is based on the version I have, so what you see may differ slightly.
Don't rush to click generate — get the body profile right first
A term first. Body means the robot itself. Action sequence is a series of joint target values over time — that's what the model outputs. Degrees of freedom: each independently rotating joint counts as one.
1. Register and log in, enter the workspace, find Body Management in the left sidebar, click New Body.
2. Fill in the number of degrees of freedom. A full humanoid is generally twenty-something to forty-something; fill in according to the unit you actually assembled, not the marketing page.
3. Fill in joint limits and zero positions. Each of these must be checked against your calibration table or URDF — angle units, signs, zero position definition.
4. Choose gripper or suction cup for the end effector; choosing wrong doesn't affect generation but affects end-effector pose solving.
5. Save, and wait for it to run a self-check. If the self-check fails it will report which joint directly.
Step three is where things most easily go wrong; I'll cover it separately below.
Write the task description in plain language
Go to Tasks, create new, select the body you just created.
Just write one sentence in the input box, e.g. "put the white block on the table into the blue box on the right." Don't write things like "execute pick-and-place task" — the model wants the correspondence between scene and goal, the more specific the better. If you have a reference image, attach one; it noticeably improves localization.
Click Generate Action. On my side a five-second pick-and-place takes roughly tens of seconds to a minute-plus; longer tasks are slower. The output is time steps plus joint targets per step, viewable as a table or directly click Replay to run it in simulation.
When replaying, mainly check for clipping, hitting limits, and whether the end effector reaches position.
Three pitfalls I hit
It ran, but not right the first time.
First was signs. When filling in joint limits I reversed the sign on the elbow joint, and the replay showed the whole arm swinging the opposite direction, looking like it was having a cramp. I'd hit this pitfall once before on joint calibration — sign errors are the most insidious; the parameters go in without error and the result is all wrong. The fix is to drag each joint individually to its limit position and check the direction, not run the whole robot at once.
Second was the task description being too vague. I first wrote "tidy up the desk" and the output just spun in place. Then I changed it to "push the black block on the left to the middle of the table" and it was right first try.
Third was long-horizon tasks. The materials mention a team training whole-body mobile manipulation with about ten thousand hours of human data — that kind of long-horizon capability is a training-level thing, not the same as you writing one sentence and it doing ten steps in a row. My approach is to break it down: split one pick-and-place into approach, grasp, lift, move, place, generate each segment separately and string them together. Tedious, but controllable.
What I'll try next
First, connect to the real robot. A clean replay in simulation doesn't mean it runs on the real machine — zero position deviation, reduction ratio, control cycle all drag down performance. I plan to take a prototype and do a small-load dry run, without vision first.
Second, run the same task on another body with a different number of degrees of freedom. Cross-body is the headline capability of this thing and the part I most want to verify — after all, an OEM can't make just one model.
I'll come back and note it down after running.
Physix Frontier