Centaur

Nine dancers taught the machine to move - then had to share the stage with what it remembered.

Centaur is a full-length contemporary dance performance in which an artificial intelligence, trained on the dancers’ own movements, becomes both performer and co-creator. Developed by choreographer Pontus Lidberg in collaboration with composer Ryoji Ikeda and AI artist Cecilie Waagner Falkenstrøm, the work explores what happens when choreography is no longer authored solely by humans, but emerges through an ongoing dialogue between bodies and machine intelligence.

Inspired by the mythical centaur - half human, half horse - the work investigates a contemporary form of hybridity: the entanglement of human creativity and artificial intelligence. Throughout the performance, dancers interact with an AI system capable of generating choreographic suggestions, poetic language, and live instructions. Each performance becomes a unique negotiation between intention and prediction, structure and emergence, creator and creation.

We needed to capture high-resolution movement data from every rehearsal. To do this we used the Sewio RTLS ultra-wideband tracking system, with sensors worn by each dancer streaming positional data in real time. A backend received that stream and uploaded it to BigQuery, where it was stored and labelled for use as training data.

Two parallel approaches were developed for generating new choreography from that data. The primary model was an LSTM with a Mixture Density Network output layer, trained to produce movement sequences from the distribution of what the dancers had already done. A genetic algorithm ran as a second compositional engine. A batch generation pipeline orchestrated both approaches. Coordinating audio timing, assembling image sequences into video, and packaging outputs. Running hundreds of choreographic variants across the studio cluster over many hours.

Live on stage, the system runs as three services communicating over OSC. The scoring service ingests the real-time sensor stream, compares it against precomputed reference files that define what each passage of the dance should look like, runs a PyTorch model to calculate dancer scores, and broadcasts those scores to the lighting and visual systems. All within the latency budget of a live performance. The stage machine sits on the local theatre network and is reachable remotely over WireGuard when the show is on tour.

Thanks to: Augustinusfonden, Overretssagfører L. Zeuthens Mindelegat, William Demant Fonden, Knud Højgaards Fond, Beckett-Fonden, Jyllands-Postens Fond