Paqarina: teaching a robot to dance Bolivian

In 14 sessions, Bolivian dancers taught an AI the tinku, the kullawada, the suri sicuri and breaking. On 20 October 2026 eight of those dances ran, standing, on a simulated Unitree G1 humanoid. On 2 November they go to a real one in Hong Kong. This is how it happened, with the people who made it.

The dancers

Everything starts with people who have danced these steps since childhood. Over fourteen sessions in Bolivia, dancers from the Ballet Folclórico and independent performers came into a white room with four cameras and danced what they know: the tinku of Norte Potosí, the kullawada of the Aymara spinners, the suri sicuri of the altiplano with its rhea-feather crown, and breaking, the one dance the young ones brought from the street.

No motion-capture suits, no markers. Each dancer wears a plain colour so the cameras can tell them apart, and the room is calibrated every morning with a printed chessboard, because a calibration from another day once cost thirteen hours of rework. Lucía danced the kullawada, Natalia and the Ballet pairs the tinku, Lu, Sean, Galgo and Keny the breaking, Veimar and Daniela the suri sicuri with its feather crown. They danced the same thirty seconds again and again, solo, in pairs, in a group, so that the machine would see the dance from many bodies and not from one.

From bodies to numbers

Four ordinary cameras, looking at the same room from four corners, are enough to rebuild a body in three dimensions. Our own software finds 26 points on each dancer in every frame of every camera, 24 times a second, and triangulates them: shoulders, elbows, wrists, hips, knees, feet, the head. A thirty-second dance becomes a table of 720 rows and 78 numbers each.

Most of the work is not the maths but the honesty of the data. A person walking behind the dancers breaks the reconstruction, so every camera is checked before triangulating. Two dancers who get close to each other swap identities in the software, so the clean stretch of a pair's take has to be measured and the rest thrown away. One frame in ten has an elbow or a wrist landed in an impossible place, so bone lengths are fixed before anything else is done. Each of these rules was learnt the expensive way, by looking at a result that was wrong and asking why.

The archive today is the whole Paqarina dataset: fourteen sessions of tinku, kullawada, suri sicuri and breaking, catalogued take by take, with who danced, in what colour, and which thirty seconds are trustworthy.

The avatar

Numbers do not dance. To see whether the capture was alive, the points are retargeted onto an avatar, designed by fashion designer Maria Corvera, built by the artist Brian Condori: a dancer in a blue jacket with a woven aguayo across the chest, bells on the boots and a long braid. In Blender the braid swings, the boots hit the floor, and a bad capture shows itself immediately as a shaking shoulder or an arm that has turned all the way round.

The hands were the hardest part. Twenty-six points end at the wrist; the palm is not in the data. For a long time the arm that holds the suri's hat twisted 360 degrees because the software was guessing where the palm faced. The fix was anatomical rather than clever: the forearm got its own twist bone, the palm was allowed to turn only within the range a human forearm can, and only while the hand is actually holding something near the head. Everywhere else the arm is left alone. In the next sessions the hands will be captured too, 21 points each.

A model that invents

Paqarina is not a motion library. The captures train a generative model, m3f, that produces new dance from the four styles. Its first full dance, a suri sicuri it had never seen performed that way, is the one that first reached the robot. Retrained on 10 October with the Ballet Folclórico sessions, the model's error on the suri dropped from 13.7 to 8.1 centimetres, measured across four random seeds rather than one, because one lucky seed proves nothing.

The second piece of thinking is a planner. Dance for a robot is a planning problem under constraints: how much floor there is, which moves can follow which, how fast a joint may turn. A small model learns to predict where a dance is going half a second and two seconds ahead; a search over its predictions chooses sequences of real two-second segments that fit the space and read as one piece. Its latest result choreographs a duet: dancer A follows a script that moves from tinku to kullawada to breaking, dancer B is planned against A's real path, kept between one and two metres away and facing her, with different material for each.

The robot

A robot cannot simply play the animation file. An avatar weighs nothing; a Unitree G1 weighs 35 kilograms, has 29 motors with limited strength, and stands on a floor. If it lifts its left leg exactly as Lucía did, its centre of mass ends up three centimetres outside the right foot and it falls; nothing in the file tells it to lean two degrees or push with the ankle. Lucía does that without thinking. The file does not contain it.

So each dance is taught, not played. The choreography becomes a reference of 29 joint angles, and a neural network learns, in Unitree's own training framework on a rented GPU in the cloud, to follow it while staying upright. It learns by falling: 4 096 simulated robots attempt the dance at once, over and over, for nine hours per dance, until the falls stop. Eight dances were trained this way between 14 and 20 October 2026.

Then comes the honest test: the policy is taken out of the simulator it learnt in and run on Unitree's official robot model, with the floor friction, motor gains, latency, mass and sensor noise varied at random, twenty times per dance. For a week the results made no sense. The calm dances stood; the fast ones fell at the same second of the same step, whatever we changed: the robot model, the physics step, the contacts, the start pose. In the end a line-by-line comparison of what the policy was being shown found it: a default pose with the knees straighter than the one it had trained with, a difference of 22 degrees at the knee, hidden in a configuration file borrowed from Unitree's own demo. With the right pose, all eight dances finished standing in 20 runs out of 20, including the jumps.

What happens next

On 2 November 2026 Violeta Ayala arrives at the CMA Lab of HKUST, where Prof. Xin Tong's group has a real Unitree G1. The eight policies were sent ahead on 20 October so the lab can run them in Unitree's own simulator first. Then the order is the careful one: the robot on a safety harness, the calmest dances first (the suri and the kullawada), then free standing, then the tinku with its jumps. Honest odds, written down before the trip: the suri and kullawada standing and recognisable on the real robot, about seven in ten; the jumping dances clean on the first try, three or four in ten; something from Bolivia dancing on a G1 in Hong Kong, better than eight in ten.

The field is moving toward the intuition everyone starts with, that a robot should simply follow the same file as the animation. In 2026 NVIDIA's SONIC and the CVPR paper Humanoid-GPT showed single networks that track dances they never trained on, on a real G1, from YouTube video. None has shown a tinku. The plan in Hong Kong is to try one of them with Paqarina's own choreographies, next to our eight policies; if it works, the nine hours of training per dance disappear and the pipeline becomes capture, retarget, robot.

And the dancers keep dancing. The next sessions add the hands, 21 points each, and the faces, recorded separately while the dancers sing, so that one day the robot's dance and the avatar's face belong to the same person.

Paqarina in numbers

People. Dance: different dancers from the Ballet Folclórico and the independent scene, across 14 sessions. Mocap system, built and run at every session: Dan Fallshaw. Avatar: Brian. Community manager: Sanam Prado, previously Rilda Paco. Direction and engineering: Violeta Ayala. Funding: Mozilla Foundation. Code: github.com/violetajaguar/paqarina.