Paqarina

Paqarina builds community-governed movement datasets with dancers across four Bolivian traditions, fusing ancestral dance with breakdancing through ch’ixi logic.

Funded by the Mozilla Foundation.

Kullawada · Weaving and spinning knowledge
June 2026 · Paqarina · Aymara Altiplano · Notebook 01
The Kullawada translates the art of spinning and weaving into a living system of knowledge: the spindle as algorithm, the thread as collective memory, the weave as community intelligence. A teaching guide with symbols, characters, costume and spatial formations, made for the dancers.
Open the notebook In Spanish.
Acta 01 · First team meeting
29 May 2026 · Paqarina · El Martadero · Log
The first meeting of the Koa.xyz lab with the Paqarina team. The project is presented, the four dances of the first phase are defined (Tinku, Suri Sikuri, Tobas, Kullawada), and the team decides to begin the creative work with the Kullawada.
Open the notebook In Spanish.

Four members of the team sit on the checkered floor of El Martadero, with notebooks, under a red roof.

Seen from one side, the circle of the first meeting on the checkered floor, with a laptop in front.

Dancers and the lab team sit along a wall and on the checkered floor, listening.

The team talks in a circle on the checkered floor, around a laptop and sheets of paper.

The archive

Tinku and kullawada have no motion dataset anywhere in the world. Paqarina builds its own, in a rehearsal room in Cochabamba, with the people who dance these dances.

4.4 hours of dance recorded so far, each take filmed by four cameras: breaking, kullawada, suri sikuri and tinku, solos and pairs. Each frame holds 26 body points. More than twenty dancers, from 17 to 70 years old. The older dancers exist in no other dataset.

Two dancers on a parquet floor, in a room with a large painting, in an early session. — June 2026, the first sessions

A breaker holds a freeze on one hand, on a white floor between camera tripods. — 13 August 2026, breaking

Three dancers dance kullawada together in a white room with tripods and lights. — 14 August 2026, kullawada, the trio

Two dancers in blue and red bodysuits dance tinku, bent forward, in a white room. — 26 August 2026, tinku

A dancer in a pink top steps across a white floor, beside a tripod. — 2 September 2026, the older dancers

A dancer in a tall feather headdress turns on one foot in a white room. — 9 September 2026, suri sikuri

A dancer in a feathered headdress leaps with open arms in a white room with lights. — 8 October 2026, Ballet Folclórico

A dancer moves across a parquet floor, beside old furniture and a large painting. — Breaking, floor work

Twelve seconds of the kullawada session, 14 August 2026, without its music.

Four ordinary cameras

Three or four cameras on tripods, calibrated with a chequerboard that every camera can see. The dancers clap on camera to sync, say the label out loud and dance two or three minutes. The films become 3D skeletons, levelled to the floor and checked bone by bone. A femur does not change length with the pose; if it does, the take goes back.

A browser app holds the archive: every take in 3D beside its camera views, with a check for every bone, tabs to calibrate, triangulate and train, and one that asks the model for new dance and sends it to Blender as a dressed avatar. Dancers and choreographers use it without touching code.

Four camera views of the same take: two dancers in a white room, seen from four corners.

The Paqarina app: the list of takes, a camera view of a dancer and the 3D skeleton of the same moment.

Two 3D skeletons in pink, cyan and white, from a duo of suri sikuri and breaking.

What the model keeps, and what it loses

An encoder learns six-second windows of movement. A diffusion model makes new windows for a given dance, and a decoder turns them back into a skeleton.

The model keeps every single quantity: speeds, heights, angles, reach. It loses the relations between them: the hip that cancels the leg, the feet that alternate, the slow phases of a planted foot. That is why generated dance slides while every number looks right.

A filter meant to remove tremor was eating the dance itself. Real tinku carries 19% of its energy above 6 Hz, in the foot strikes and the hand accents.

Every claim carries its measurement. Conclusions we take back stay in the record. Nothing is believed from a single random seed.

A chart of the dances in the space the model learned: breaking in pink, kullawada in cyan, tinku in yellow.

Charts of two dancers of a trio: the distance between them, how much each moves, and how their movements correlate in time.

The avatar and the robot

The real and the generated dance drive an avatar in Blender and a Unitree G1 humanoid in MuJoCo simulation. A balance policy learned to dance a whole 30-second tinku phrase standing, with no harness, from 12 of 17 starting poses. What unlocked it was taking the harness away in training: the policy had learned to hang from it.

What the avatar does best, the most vigorous tinku in the archive, is exactly what the robot cannot do.