Where you are. You have opened three architectures whose papers or weights you can read. This one you cannot, and that is the lesson. It is also the clearest illustration in the field of why the fast-slow split exists at all.
Four minutes, no cuts, and the tab you had open in week one
You have watched this video already. It is the one in the left-hand tab of lesson 0.8, where it was the example of a demonstration that supports exactly the claim it makes. Open it again, now that you know what is inside the machine.
A humanoid robot empties a dishwasher. It walks to the machine, opens it, picks up a plate, turns, walks to the cupboard, puts the plate away, comes back. It does this for four minutes. Sixty-one separate actions, one after another, no teleoperator, no human stepping in to reset the scene, no visible edit in the middle of the sequence.
Your reaction to that video is correct and you should keep it. Nothing you build in this course will do that, and nothing openly available will either. It is genuinely further ahead than the things you can download.
Now ask the second question, the one the video is not designed to prompt: how many times did they film it? That number is not published. It is not published for any robot demonstration video, by anyone, ever. And without it, the four minutes tells you something real but much narrower than it feels.
The idea in one paragraph
Figure’s Helix is, on the evidence available in mid-2026, the most impressive humanoid manipulation system anyone has shown in public. It also has no paper, no benchmark, no evaluation protocol, no ablation, and no weights. Its architecture is worth learning anyway, because it draws the fast-slow line in three places rather than two, and the three rates it runs at make the argument from clock domains impossible to miss. But learning a design and believing a claim are separate acts, and this lesson is about keeping them separate. By the end you will have a three-bucket test you can apply to anything: downloadable, published, demonstrated, and a clear statement of what each one entitles you to conclude.
Three clocks stacked
Figure describes Helix 02, announced 27 January 2026, as three systems.
Wider than the screen; scroll it sideways.
System 2 is the slow half you have met three times now: semantic reasoning, scene understanding, language. It emits a latent goal. Figure does not publish its rate, but “slow” here means the tempo at which a task changes, roughly once a second, not the tempo at which a wrist moves.
System 1 is a visuomotor transformer running at 200 Hz, conditioned on System 2’s output. It consumes head cameras, palm cameras, tactile sensing and proprioception, and emits full-body motor commands. Note the input list: this is the first design in the module where touch is part of the loop, and palm cameras exist for exactly the reason your own hand occludes what it is grasping.
System 0 is new in Helix 02 and is the interesting one. It is a whole-body controller of roughly 10 million parameters running at 1 kHz. It was trained entirely in simulation, across more than 200,000 parallel environments with domain randomisation, on more than 1,000 hours of human motion retargeted to the robot’s joints.
Why three tiers and not two
Read the rates on the diagram again: slow at the top, 200 Hz in the middle, 1 kHz at the bottom.
There is no design fashion in that ladder. It is physics, three times over. Understanding a sentence has no physical deadline. Reaching for a plate has a deadline set by how fast the plate and the hand move relative to each other, which is tens of milliseconds. Not falling over has a deadline set by how fast a body topples, which is milliseconds. A humanoid needs a tier a table-top arm does not, because a humanoid can fall.
This is the cleanest evidence in the field for the claim lesson 4 made abstractly: these layers are separate because they are different clock domains, and their rates come from the world rather than from the model designer.
What the demonstrations claim
The dishwasher run is the headline: four minutes, 61 sequential locomotion-and-manipulation actions, autonomous, no resets. Alongside it Figure shows unscrewing a bottle cap, extracting a pill, and dispensing 5 millilitres from a syringe. That last set is dexterity of a kind nothing in this module’s downloadable tier comes close to.
Here is the honest read. Each of those is a claim about a specific take. The number of attempts is not disclosed, the failures are not shown, the videos are edited, and there is no protocol against which anyone could try to reproduce them. None of that means the demonstrations are fake. It means they carry a different kind of information from a benchmark: they are an existence proof that the capability is reachable, under conditions you were not told.
The three buckets
Almost all the confusion in this field dissolves when you sort claims into three buckets before reasoning about them.
Wider than the screen; scroll it sideways.
The buckets are not a ranking of quality. Helix is almost certainly more capable than anything in the first column. They are a ranking of what you are entitled to conclude, and the two rankings are close to unrelated.
A downloadable model lets you find out whether it works for you, which is the only question that ever actually mattered. A published model tells you an idea is real and will be replicated or refuted by someone. A demonstrated model tells you a capability exists somewhere in the world, which is genuinely useful for deciding what to attempt and useless for deciding what to deploy.
Why the frontier stays closed
Notice what is public. Every architecture in this module came from public sources: two experts in one attention stack, a backbone feeding a diffusion head, two models passing a sentence, three tiers at three rates. Companies publish their architectures freely, and the diagrams in this module are the proof.
What nobody publishes is the corpus. GR00T’s 20,000 hours of egocentric human video, Physical Intelligence’s training mixture, Figure’s fleet hours: none of it is available, and it is the part that took the money and the years to collect.
Wider than the screen; scroll it sideways.
That is the actual shape of the moat, and it has a direct consequence for you. You can copy any of these designs. You cannot copy the data. So the leverage available to one person with one cheap arm is not architecture and is not scale; it is your own data on your own robot, applied to a pretrained model somebody else paid to build. Which is precisely what the rest of this module does, starting with the one you can run.
Check yourself
1. Why does Helix need a tier that a table-top arm does not?
Because it can fall over. Balance is an unstable equilibrium with a time constant measured in milliseconds, so it needs a controller sampling at around 1 kHz. A bolted-down arm’s fastest deadline comes from contact rather than from stability, which is demanding but an order of magnitude slower. The extra tier is a consequence of the body, not of the intelligence.
2. System 0 has about 10 million parameters and runs a thousand times a second, while the layer above it is hundreds of times larger and runs far less often. Why is that the right way round?
Because capability and deadline trade against each other. At one millisecond per decision there is no compute budget for a large model, and there is also nothing useful for one to do: keeping a body upright is a narrow, well-posed control problem that a small network trained on enormous amounts of simulated motion can solve. The semantics that need a large model do not need to happen every millisecond.
3. A four-minute video shows 61 actions with no resets. What exactly does it license you to conclude?
That the system completed that sequence at least once, under conditions that were not disclosed, on a robot you cannot buy, with an unknown number of prior attempts. It is an existence proof of a capability, which is real information. It is not a success rate, and treating it as one is the single most common error in reading this field.
4. Why is Figure’s logistics data-scaling result more credible than its dishwasher video, even though both come from the same company with no protocol?
Because of its shape. It reports two endpoints on a curve, in specific units, for a quantity that further deployment could contradict, and it makes a general claim about how performance responds to data rather than showing one impressive outcome. Neither has an independent check, so credibility here is about falsifiability, not verification.
5. Every lab publishes its architecture and none publishes its training corpus. What does that tell you about where the value sits, and what should you do about it?
It tells you the architecture is not the moat: it is cheap to describe, largely converged, and copying it buys you nothing on its own. The corpus is the expensive, unreproducible asset. The move available to an individual is not to compete on either but to bring data nobody else has, from your own robot and your own task, to a pretrained model somebody else funded.
6. Which bucket does a model sit in when the weights exist but require an application form?
Not the first one. An application is a gate, and a gated model is an announcement until the day it is in your cache. Sort by what you can obtain right now without asking permission, because a plan built on a pending application is a plan with an unbounded blocking call in it.
Do this
About thirty minutes, all reading and writing.
1. Bucket the claims. Watch Figure’s Helix 02 announcement video. For each distinct capability claimed, write one line: the claim, the bucket it sits in, and the single piece of missing information that would move it up a bucket. Most lines will end with the same answer, which is the finding.
2. Draw the ladder. From a blank page, draw the three tiers: name, rate, what it consumes, what it emits. Then mark the one arrow you are least confident about and write why. Checking your uncertainty is as much of the exercise as checking your recall.
3. Run the pass on something new. Find any robot demonstration posted in the last month, from any company. Apply the same three questions: what bucket, how many takes were disclosed, and what would you need to see before planning around it. Do this once and it becomes automatic, which is the actual deliverable of this lesson and the reason reading the frontier exists.
What you can now do
You can draw Helix’s three-tier design from memory and explain why a humanoid needs a rung a table-top arm does not. You can sort any robot claim into downloadable, published or demonstrated, and state precisely what each bucket entitles you to conclude without arguing about whether a demonstration is real. You can tell a claim shaped like evidence from one shaped like a highlight reel. And you know where the frontier’s advantage actually sits, which tells you where yours does not: not in the architecture, which is public, but in data, which is the one thing you can go and collect.