Edgar 🦆

World Models.Embodied AI.

the future it keeps
futures it imagined and dropped

How a world model thinks

1Learn the bodyOnce. Task-free motion, no reward, no goal. A JEPA learns to predict the body in a tiny latent space, then it is frozen.
every 50 ms
2ImagineHundreds of possible futures, rolled through the frozen model.
3JudgeThe task is a cost. Each future gets a score.
4ActBlend the best ones. Execute the first step only.
5FeelMeasure what really happened. Events switch phases, never a clock.

The same loop flies a plane, carries a person, shakes a cocktail. Only the body's model and the task's costs change. Built on I-JEPA · V-JEPA 2 · LeJEPA · MPPI.

The world is not
one kind of data

VisionKinematics + physicsSensors + electricityMathematicsSafetyComputeLanguageWho owns the weights
One modelbillions of weights · every nature mixed into one token stream
Visionthe mapwhat is where, remembered
Kinematics + physicsthe body's world model5k–30k parameters, frozen
Sensors + electricitythe nervesservo load, line tension, motor current
Mathematicsthe plannerMPPI: futures, costs, weights
Safetycosts you can reada rule is a line of code, not a hope
Computethe meterevery millisecond of every tick
Languageonce per sentencewords become skills, never in the loop
Who owns the weightssmall enough to ownread, audit, run on your own chip
one live loop · each nature in the model that fits it

Small models.
Live thinking.

4.1B28kparameters: the VLA route · ours
97 h0demonstrations: teleoperated · ours
350 H100·h11 mintraining: GPUs · one laptop CPU

Same cable robot, same job. Theirs is the published route on the real machine; ours is simulation so far.

Where it started

June: the first arm

50 teleoperated demos for one task

A policy learned from them picked a rubber up, on some positions. Fifty demos per task, per object, per room: it does not scale, and it does not generalise.

July: SkyFall

0 demos · a 5.6k world model · 3 drops between pillars

Learn the drone once, task-free. Every mission is then a cost: fly between the pillars, drop each ball within ~0.23 m.

October: every body

the same recipe, carried from body to body

Quads, a plane, two arms with eyes, a cable robot. One loop, one planner, a small model per body.

Drones

A quad lands on a plane in flight

two world models, each planning around the other's future

Quadrotors in a forest

5.6k parameters · SkyJEPA

Plane

5.5k parameters · 2–3× tighter than physics

Through a canyon

terrain as a cost · hold, fly, drop

Arms and eyes

The real arm

real SO-101 · grasp 6.4 s · in the cup 15.4 s

Pen in the cup, 24 times

17 / 24 random layouts · ~10 s · one wrist eye

One eye, two eyes

7 / 7 things kept · 0 phantoms · 58 ms a look

A head that sees in 3D

1.7 mm median error on the surfaces

Next: 2 arms, 2 eyes, one body

The bust

printed trunk · owl neck · stereo head · two SO-101

Its eyes at work

two head eyes, two arms, one map

Robots together

Three world models, one cigar

arm → drone → arm → Tony, and back home · every hand-over decided by touch · turn the sound on

Hand-over by feel

pen, rod, cylinder · 4–9 s each · let go when the taker holds

Two arms, one shaker

3 / 3 · 0 drops · the bartender's stroke

Throw onto the mark

within 30 mm · release timed by the hand's curve

For people.
For industries.

Autonomy after paralysis

bed to chair · 50 to 150 kg · seated every time

Avalanche rescue

three drones · the point found 0.18 m from the truth

Home help

one camera · 6 / 6 things kept · ~2 mm

Kitchens and bars

stirring by feel
A scale-accurate twin of a winery, from the vines to the tank hall

Factories and vineyards

a winery's twin, vines to tanks

Festivals

humanoids serving drinks, funny with the crowd

Drones carry the diggers

drones bearing robots to the avalanche

Defense

offensive and defensive robot teams

If these were billions

5.5k
a plane's body · today
28k
a cable robot, any room · today
10M
two arms and all they handle: the hand-over learned, not written
1B
a house: every door, drawer and machine, and how each behaves
10B
a factory or a vineyard: machines, crops, seasons, routines
100B
a city's robots sharing one imagination

Today's models are tiny on purpose: one body each, small enough to read and to run on a chip. Scaled like the language models, the same recipe would not grow a bigger chatbot. It would grow a bigger body: more things it can touch, imagine and keep safe.

papers quad lands on a plane plane model cable robot model SkyJEPA GitHub