Velkorti

robots • hardware • humanoids
What the demo didn't show.

[Velkorti @velkorti]

Every humanoid robot demo you watched this year was solving a problem that was answered in 1990 by a pair of legs with no motors in them.

The course that explains why is free. MIT charges $92,760 a year for the room it is taught in.

His name is Russ Tedrake. He teaches 6.832, Underactuated Robotics, and he also runs robotics research at Toyota Research Institute. Lectures on YouTube, textbook on a public web page, code on GitHub.

The whole thing turns on one word.

A robot arm bolted to a floor is fully actuated. A motor for every joint, so you can command it into any pose you like. A robot that walks is not. It is attached to nothing, and gravity is doing half the work. Fewer motors than the motion actually requires. That is underactuation, and it is why walking is a different problem from lifting, not a harder version of it.

The expensive answer is to add motors and fight the dynamics. Honda's ASIMO had one for nearly every joint and still took stairs like a man carrying soup.

The legs were the other answer. An engineer named Tad McGeer built them, set them on a shallow slope and let go. No motors, no sensors, no computer of any kind. Gravity and the swing of their own limbs did the entire job.

Tedrake opens the course with that machine sitting next to ASIMO and lets the comparison do the talking.

After that you cannot unsee it in the footage. The robots that move well are borrowing from their own weight. The ones that look wrong are overpowering it, and you can hear it in the servos.

The course is free. McGeer's paper is thirty-six years old.

Tedrake still teaches it at MIT. Almost nobody reposting robot videos has watched a single lecture of it.

The theory is free. Knowing which video you are actually looking at is the whole edge.

[Velkorti @velkorti]

@DigitalDaisyX Small detail: that's an AgiBot, not Optimus

[Velkorti @velkorti]

@ZentrexQ The cowboy hat is doing a lot of the trust work here

[Velkorti @velkorti]

There won't be a day when a robot sends 500 workers home.

Instead:
A night shift stops hiring.
A warehouse doubles with the same team.
A plant ships 30% more with zero job ads.

Nobody will notice. That's the point.

[Velkorti @velkorti]

There won't be a day when a robot sends 500 workers home.

Instead:
A night shift stops hiring.
A warehouse doubles with the same team.
A plant ships 30% more with zero job ads.

Nobody will notice. That's the point. https://t.co/1QbQCxaT2X

[Velkorti @velkorti]

There won't be a day when a robot sends 500 workers home.

Instead:
A night shift stops hiring.
A warehouse doubles with the same team.
A plant ships 30% more with zero job ads.

Nobody will notice. That's the point. https://t.co/1QbQCxaT2X

[Velkorti @velkorti]

@0xBynode Well, at least he tried to do it perfectly 😅

[Velkorti @velkorti]

@itchson That's very sweet.

[Velkorti @velkorti]

@LeoKharon Zero embodiment gap is a big claim. Would love to see it on a messy task

[Velkorti @velkorti]

A $100,000 HUMANOID LIFTS 200 KG HALFWAY AND THE TIRE DECIDES THE REST

Not because the tire is heavy.
Because its weight moves while it's lifted.

Watch the failure chain:
hands catch the rubber edge at the wrong angle
tire tips toward the robot, not away
knees take a load the plan never had

Lifting a box is one force in one direction. A flipping tire moves its centre every centimetre, and the body has to chase it.

Lab demos sell a robot that lifts 20 kg.
The quarry asks what happens when the weight fights back.

[Velkorti @velkorti]

@_robtimus_ The rope says more about trust than the robot does

[Velkorti @velkorti]

A $150,000 HUMANOID HAS ONE STEP TO STOP ON A 30 CM LEDGE OR KEEP FALLING

Not because the gap is wide.
Because landing is the jump nobody trains.

Watch the failure chain:
feet hit the ledge with forward momentum left over
ankles lock to stop it
upper body keeps going over the edge

Clearing the gap is physics. Stopping on the other side is control: bleeding off speed in one contact, with no second step available.

Lab demos sell a robot that jumps across.
The rooftop asks whether it can stay there.

[Velkorti @velkorti]

@aistasiia 69% still demos and labs. Real work is the other 31%

[Velkorti @velkorti]

@Sanjil Robots building robots. The factory is about to get very quiet

[Velkorti @velkorti]

@melrate the boring one on wheels. it can't fall on me

[Velkorti @velkorti]

@CyberRobooo rain is the real test, not the demo stage

[Velkorti @velkorti]

A $120,000 HUMANOID LOSES THE BALL IN UNDER ONE SECOND ON A CRACKED COURT

Not because the ball is fast.
Because the ground decides where it bounces.

Watch the failure chain:
ball hits a crack and kicks sideways
hand arrives where the ball should have been
second touch pushes it out of reach

Dribbling is catching something you just threw at the floor, every half second, without looking. The asphalt edits every throw.

Lab demos sell a robot that catches a ball tossed to it.
The court asks whether it can catch one the ground threw.

Stress test #16. The rest are on my page.

[Velkorti @velkorti]

@GradientX0 Carrying groceries is the job that actually keeps older people living at home longer. This matters.

[Velkorti @velkorti]

@bot1kai From falling on a door handle to tying surgical knots in two years. That's the real progress curve.

[Velkorti @velkorti]

@CyberRobooo Panasonic already makes the motors and batteries. Building the whole robot was only a matter of time.

[Velkorti @velkorti]

A $70,000 HUMANOID GETS ONE TIMING ERROR OF 50 MILLISECONDS AND THE ROPE WINS

Not because jumping is hard.
Because the rope never waits.

Watch the failure chain:
wrists turn a fraction late
jump launches on the old timing
rope meets the shins on the way down

Every skip is two systems that have to agree: hands that swing and legs that leave the ground. Drift by a hair and they stop agreeing.

Lab demos sell a robot that jumps once.
The rope asks whether it can jump the 200th time like the first.

Stress test #15. The rest are on my page.

[Velkorti @velkorti]

@0xNextCore Animals are the hardest coworkers for a robot. They don't follow routines, and every animal reacts to a machine differently.