Research

Paqarina generates the dances it learned, for an avatar and for a robot

Paqarina, a Mozilla Fellowship project, learns the dances of our community from the bodies of the people who dance them. Then it generates them: as novel dance that nobody danced, or as a remix of what was danced, joined by how the pieces fit together. And because it is modular, the same dance moves two bodies: an avatar in Blender and a humanoid robot.

The process

The process, from the dancers to an avatar and a robot. Video: Paqarina.

It learns from dancers

Since June 2026 dancers of our community have danced for four cameras. Tinku. Kullawada. Suri sicuri. Breaking.

No suits and no markers. From the four views we rebuild each dancer as 26 points in space, 24 times a second.

On 8 October six dancers of the Ballet Folclórico danced for us: Lucía, Valeria, Juan David, Natalia, Veimar and Paolo. Lu, Sean and Galgo dance breaking.

What Paqarina keeps is the dancing itself, cut clean: 137 takes of our own, 4.4 hours of movement. Its model also learned from 101 short takes of street dance from a public collection.

It generates dance

Paqarina generates every dance it has learned. It does it in two ways.

Novel dance. The model starts from noise and writes movement that nobody danced, in the dance it is asked for. The suri sicuri below is one. No person danced it. Measured against real suri, it stays 8.1 centimetres per point away.

A novel suri sicuri, written by Paqarina's AI model, on the avatar and on the robot's model. The robot here only copies the angles: it is not balancing yet. Render: Paqarina.

Remix. Paqarina takes pieces of what was danced and joins them according to how they fit together: each piece begins in a pose close to the one where the piece before ended. It can be asked for an order of dances, for a stage to stay inside, for a dance that stays on its feet. Tinku 10, kullawada 10, breaking 10 is a score it can read.

Either way, what comes out has the same form as what went in: 26 points, 24 times a second.

One dance, two bodies

Paqarina is modular. The 26 points are where its pieces meet. On one side is where the dance comes from. On the other are the bodies that dance it. Any dance can go to any body, and one piece can be changed without touching the others.

PieceWhat it readsWhat it writes
CaptureFour camera views of a dancer26 points
Novel danceWhat the model learned from 238 takes26 points
RemixPieces of takes, and an order or a score in words26 points
The avatar26 pointsA Blender file with the dance in keyframes
The robot26 pointsA table of 29 motor angles
TrainingThe table of anglesA network that keeps the robot on its feet

The avatar in Blender

The avatar has its own bones, its own proportions and the costume.

Each of its bones points at the dancer's next joint: the upper arm at the elbow, the forearm at the wrist, the thigh at the knee. So the avatar takes the directions of the dance and keeps its own lengths.

Some things are not in 26 points. They end at the wrist, so the palm is not in the data: a rule turns it toward the body. The wrist is kept out of a sphere around the head, so a hand does not go through the face. Recorded feet are trusted as they were measured. The braids take a colour for each dance: here, green is kullawada.

What comes out is a Blender file with the dance in keyframes. It opens and plays without our program, and it can go on to other 3D tools.

The same eight seconds of kullawada on two bodies. Left: the avatar in Blender. Right: the robot's model holding its 29 angles. The same 26 points move both. Render and simulation: Paqarina.

The robot

The robot is a Unitree G1. It is 1.32 metres tall, weighs about 35 kilos and has 29 motors.

It is not a dancer made smaller. From shoulder to wrist Lucía measures 44 centimetres in our data and the robot 38. From hip to ankle she measures 71 and the robot 64. So every segment of the body is brought to the robot's length on its own, keeping the direction it points in.

Then, for every instant, we look for the pose that brings 20 points of the robot closest to those points: 14 joints and 6 points of the feet. Every angle stays inside what its motor can turn. For the kullawada Lucía danced, the robot ends up 2.4 centimetres from them, on average.

One instant of a kullawada in three steps. Left: Lucía as four cameras recorded her. Centre: the same points with every limb at the robot's length. Right: the robot's model holding the 29 angles that bring it closest. Images: Paqarina.
One instant of a kullawada in three steps. Left: Lucía as four cameras recorded her. Centre: the same points with every limb at the robot's length. Right: the robot's model holding the 29 angles that bring it closest. Images: Paqarina.

The feet get special care. When a foot is low and still in the recording, it is held on the floor. In the kullawada that is 35 per cent of the time. A sliding foot would give the robot two orders at once: stand on this foot, and move it.

What comes out is a table: where the pelvis is, how it is turned, and 29 motor angles, 30 times a second. With it comes a report against the robot's limits. In the kullawada the two ankles sit at the end of their sideways tilt, 15 degrees, for about a third of the dance. The report does not decide whether the robot can dance it. It says where to look.

Staying up

The avatar can stop there. It has no weight.

The robot cannot. A robot that only copies angles falls, because nothing in an angle keeps a body on its feet. So each dance is practised in simulation until a policy can hold it, and then it is checked in a second simulator. All eight dances stand there.

What we ask the robot to do

Everything we ask the robot to do comes out of Paqarina. Of the eight dances the robot has practised, one is novel, four are remixes, and three are single takes that Paqarina cut to their liveliest thirty seconds.

A dance tells each motor what angle to hold, 50 times a second. Four of the eight, measured:

What is askedA novel suriTinku, then kullawadaTinku, kullawada, toprockThe kullawada Lucía danced
How it was madeNovel, written by the modelRemix, 2 pieces of 15 sRemix, 3 pieces of 10 sOne take
Length21 s32 s31 s32 s
Motor angles sent31,08845,96545,58846,313
Farthest from where it starts98 cm148 cm122 cm133 cm
Turning of the hips, added up797 degrees2,128 degrees2,449 degrees1,911 degrees
Highest foot53 cm60 cm55 cm25 cm
Deepest knee bend128 degrees134 degrees128 degrees103 degrees
Highest hand144 cm144 cm161 cm134 cm

The novel suri travels least and turns least. The remixes turn the most. The kullawada keeps its feet closest to the floor, which is why it was the first dance we slowed down for the real robot.

On the real robot

In November the eight dances went to a real Unitree G1 at the Hong Kong University of Science and Technology, on its Guangzhou campus. On 5 November a kullawada took the robot's motors and held for six seconds.

Embodied intelligence

This work is about intelligence that lives in a body.

It begins with memory. Paqarina keeps the dances of our community as movement, not as description: every step danced for the cameras is there as 26 points in space.

It becomes generation. From that memory Paqarina writes novel dance and remixes what was danced.

And it has to be carried by a body. The avatar wears the costume and shows the dance. The robot has to stand it: the network keeps 35 kilos upright while it dances, and it learns that with a body of its own, by falling thousands of times in simulation.

The dances belong to our community. Paqarina generates them. The avatar shows them. The robot is learning them.

Thanks

Paqarina is a Mozilla Fellowship project.

Thanks to Dr Pan Hui, Professor Xin Tong and Jie Yu for allowing us to test our policies on the robot.

What the record shows

MeasureValue
Points kept for each instant of dance26
Dancing of our own that Paqarina learned from137 takes, 4.4 hours of movement, recorded since June 2026
Takes the model learned from, in all238: 137 of our own and 101 from a public collection
Ways Paqarina generates dance2: novel dance and remix
A novel suri, distance from real suri8.1 cm per point
Bodies the same 26 points move2: an avatar in Blender and a Unitree G1 robot
Dances the robot has practised8: 1 novel, 4 remixes, 3 single takes
Shoulder to wrist, Lucía and the robot44 and 38 cm
Hip to ankle, Lucía and the robot71 and 64 cm
Fit of the kullawada Lucía danced on the robot2.4 cm per point
On the real robot, 5 November6 seconds of a kullawada

What remains open

Whether Paqarina can learn what the robot's joints allow, so that what it generates already fits them. Whether the same 26 points reach a second robot. Whether the dancers recognise their dances in what Paqarina generates. On the avatar every dance plays whole. On the robot a kullawada has held for six seconds. The next step is to let what Paqarina generates know the robot's body.