Movement that outlives its first question
How Global Movement Research became a computational system for studying human movement over time
Movement that outlives its first question
Global Movement Research began with a question I have been working through for more than a decade: can human movement be preserved without freezing it as an archive?
Dance has never lacked recording technologies. We have notation, photography, film, video, motion capture, and increasingly sophisticated sensing systems. Each can preserve something important about a movement event. The difficulty begins afterward. A recording is usually organized around the study, performance, or collection that produced it. Once that purpose has been served, the material may remain stored, but its capacity to participate in new research is limited.
I wanted movement to have a longer life than that.
That question eventually led to the Movement Language Model, or MLM. The MLM gives captured movement a computational representation that allows it to remain searchable, comparable, and available for new analysis as the corpus grows. A recording produced for one research question can later be examined through another method, compared with movement that had not yet been captured when the original study took place, or placed inside a larger structural question across traditions, practitioners, and environments.
The recording survives its first question.
From recording movement to representing it
A conventional movement study begins with a body and an instrument. Motion capture produces coordinates. A force plate records ground reaction force. Electromyography records muscular activation. A camera preserves an audiovisual event. These systems give researchers precise evidence about a defined aspect of movement.
Global Movement Research keeps those original records, but adds another layer.
Movement entering the MLM is transformed into a shared computational representation through which recordings produced at different times, through different systems, and by different bodies can be brought into relation. The source material remains intact. The representation gives the system another way to work across it.
A moving body generates several forms of evidence at once. Skeletal movement, sound, pressure, effort, environment, speech, objects, facial movement, and practitioner interpretation can belong to the same event. They are not interchangeable signals, and the research does not treat them as though they are. Each requires its own method of capture and analysis. The MLM provides a computational structure through which those different records can eventually be queried together.
The movement codebook
The current MLM begins with skeletal movement.
Raw motion is normalized into a canonical skeletal representation and converted into movement features. The system then learns a codebook of recurring computational units across the corpus. Continuous movement becomes an ordered sequence of discrete codes.
This gives the corpus a structure that continuous motion alone does not easily provide. Sequences can be aligned, searched, clustered, counted, and compared. We can examine where structures recur, how their ordering changes, which codes repeatedly appear together, and where recordings occupy related regions of the learned movement space.
The learned codes are statistical units. They are not kinemes.
The kineme vocabulary comes from a different research process involving ethnochoreological analysis, practitioner knowledge, cultural research, and comparison across movement traditions. Kinemes describe movement through distinctions that have been established as meaningful within the research. A learned code may correspond to part of a kineme, several variations of one kineme, a transition between kinemes, or a structure for which no semantic unit has yet been characterized. A kineme may also require several learned codes.
Keeping those two vocabularies separate allows us to ask whether the computational representation preserves distinctions that matter within movement practice rather than assuming that statistical recurrence and cultural significance are the same thing.
Cultural conditioning
Dance made another problem visible very quickly. A skeleton can tell us where the body moved, but it cannot explain the full conditions through which that movement developed.
African and African diaspora traditions in the corpus contain complex relationships among weight, rhythm, articulation, environment, sound, social function, geography, and transmission. Two sequences can appear structurally similar while belonging to very different practices. Conversely, movements that look different at first may share deeper organizational structures.
For this reason, the MLM is being developed with cultural conditioning alongside movement representation.
A recording enters the research with information about its practitioner, provenance, geographic location, cultural context, apparatus, permissions, and other conditions relevant to the study. These do not become substitutes for movement analysis. They give researchers additional evidence for investigating the relationships the computation surfaces.
If two recordings sit close together in the learned representation, the model has found a structural relationship. It has not established why that relationship exists. Historical transmission, shared environmental conditions, common biomechanical solutions, related practices, or statistical coincidence may all produce resemblance. Determining which explanation is plausible requires evidence beyond the model.
What the MLM can return
The MLM is designed as a computational engine that other parts of SEE Lab can call.
At its current and developing stages, it supports or is being built to support:
- retrieval of structurally related recordings;
- comparison of sequences through time;
- inspection of learned-code patterns;
- alignment between recordings;
- comparison between learned codes and kineme analysis;
- conditioned queries across information such as geography and tradition;
- clustering and similarity analysis;
- and Movement Genome profiles.
The result may be a recording, an aligned sequence, a similarity score, a cluster, a recurring token structure, or another computational relationship that warrants closer study.
This is where the model becomes useful as research infrastructure. It does not need to know which studio will eventually display the result or how a researcher will visualize it. A studio asks the MLM for a particular reading, and the model returns structured information that the studio can use.
One movement, several analytical methods
The same movement can also be examined through several methods without forcing those methods to agree.
Qualitative Trajectory Calculus reads spatial relationships between tracked points. Sequence Alignment Methods compare the temporal organization of movement phrases. Kineme analysis reads culturally established movement units. Learned-code analysis examines statistical structures produced through model training.
Each answers a different question.
QTC may reveal that two tracked points approach and separate through the same pattern. Sequence alignment may show that two phrases follow similar temporal structures. Kineme analysis may identify a distinction that neither method captures. The learned codebook may surface a recurring structure for which the existing vocabulary has no name.
The value is not in collapsing these readings into one result. It is in being able to examine where they agree, where they diverge, and what those differences reveal about the movement or the analytical method itself.
The MLM within Global Movement Research
The Movement Language Model is one component of a larger research system.
Global Movement Research follows movement through the Dance Replication Process, a nine-stage process:
- Discovery
- Dispersal
- Analysis
- Replication
- Data
- Interpretation
- Contextualization
- Visualization
- Dissemination
The MLM is the engine behind the whole Process, not a stage inside it. It runs from the first stage: Discovery already calls it to bind a recording to the corpus, Analysis calls it to compare, and at Data movement becomes computationally comparable in its terms. Interpretation, Contextualization and Visualization read what it returns.
The distinction between the two is still important. The DRP organizes the research journey. The MLM is the computation the journey runs on, at every stage.
Contextualization is where computational findings are examined alongside historical research, geographic evidence, practitioner knowledge, social function, and other relevant sources. Dissemination addresses how the resulting research can circulate and under what conditions.
The model does not perform the entire research process, and it was never intended to.
A shared engine across SEE Lab
The same architecture allows different SEE Lab environments to work with the corpus without requiring a new model for every research question.
A studio may request movement related to a particular sequence. Another may request records conditioned by geographic information. Another may ask for an artifact representation to be compared with movement already in the corpus. An immersive environment may request several related recordings so that they can be placed together spatially.
The MLM does not become those studios. It supplies the common computational layer underneath them.
This also means movement does not have to be computationally rebuilt every time it enters another part of SEE Lab. The same underlying representation can support different forms of inquiry while the original recordings remain available as evidence.
Why this matters for the archive
Human movement carries knowledge through teaching, repetition, ceremony, migration, adaptation, and practice. Recording has allowed fragments of that knowledge to survive, but preservation has usually meant keeping the record rather than keeping the movement computationally active.
The MLM proposes another possibility.
A movement recorded today can remain available for comparison with material captured years later. A relationship that was invisible when the recording entered the corpus may become visible as the corpus expands. A method developed in the future may be applied to material captured before that method existed. A researcher can revisit an old recording without treating the original study as the final question the movement is allowed to answer.
That is what I mean by persistence.
The archive remains an archive. The recordings remain records of particular events. What changes is that they also become part of a computational system capable of forming new relationships across what it holds.
What is built now
The current system includes the skeletal movement pipeline, normalized representation, tokenizer, trained codebook, and the developing kineme vocabulary. These have been built and tested against the committed corpus.
The next stage is bringing the served Movement Language Model fully online so that corpus-wide retrieval, conditioned querying, and broader cross-cultural comparison can operate through the platform.
The original recordings and sensor streams remain the source evidence. The MLM adds a computational layer through which that material can continue to be studied as new movement, new methods, and new research questions enter the system.
That is the central proposition of the Movement Language Model: capture should not be the end of a movement record’s research life.