The Movement Alphabet That's Teaching AI How Bodies Actually Feel

Laban notation was invented in the 1920s to describe what no camera could capture. A century later, AI researchers are rediscovering why it matters.

Rudolf Laban was watching a problem that hasn't gone away.

It was the 1920s, and Laban — a Hungarian dancer, choreographer, and movement theorist — was trying to find a way to write down dance. Not just the shapes the body made, but the quality of how it moved. The effort. The weight. The sustained pressure of a descent versus the sudden lightness of a rebound. The difference between a movement that reaches toward the world and one that retreats from it.

Standard music notation could capture rhythm and pitch. Visual sketches could capture the outline of a shape. But neither could capture what dancers actually care about: not where the body is, but how it got there, and what it felt like to arrive.

Laban's solution was to build a notation system from the ground up — one that would encode movement quality, not just movement shape. The result was Labanotation (and its companion framework, Laban Movement Analysis), a symbolic vocabulary that can describe whether a gesture is sustained or sudden, strong or light, direct or indirect, bound or free. It encodes effort. It encodes spatial intention. It encodes the difference between a movement that is merely executed and one that is meaningfully inhabited.

For most of the twentieth century, Labanotation was used by dance archivists and somatic educators. Most people in the broader world had never heard of it.

AI researchers are changing that.

What Laban Notation Actually Encodes

Before explaining why AI researchers care about Labanotation, it helps to understand what it actually describes — because it's more precise than the word "quality" might suggest.

Laban Movement Analysis organises movement description around four main categories, known by the acronym BESS:

Body — which body parts are moving, in what sequence, and in what relationship to each other. (The torso leads; the arms follow; the head resists.)

Effort — the dynamic qualities of movement along four "motion factors": Weight (strong vs. light), Space (direct vs. indirect), Time (sudden vs. sustained), and Flow (bound vs. free). Each motion factor is a continuum, not a binary — a movement can be very slightly bound, or very strongly free. Effort is the dimension that most directly corresponds to the felt quality of movement.

Shape — how the body changes its form in relation to itself and to space. Does the shape spread or enclose? Rise or sink? Advance or retreat? Shape captures the mover's spatial relationship with the environment.

Space — the directional and spatial patterns of movement in the kinesphere (the personal space surrounding the body) and general space. Does the movement trace a straight line, a curved arc, a complex three-dimensional pathway?

Taken together, these four dimensions describe a movement in terms that a dancer or somatic practitioner would recognise as the actual content of their work. Not "the left arm moved to position X at time T" but "a sustained, strong, gathering movement that began in the extremity and drew inward toward the body's centre."

The Problem AI Motion Generation Has Always Had

AI motion generation systems — the kind that produce realistic animations of human movement from text descriptions or audio — have historically been trained on data that encodes only a small fraction of what Laban notation captures.

Standard motion capture data records joint positions over time: 3D coordinates for each joint at each frame. From this, you can reconstruct a skeleton moving through space. What you cannot reconstruct is the effort quality — whether that movement was light or strong, sudden or sustained, free or bound. The sensor only captures geometry. The qualitative texture of the movement is invisible to it.

Training an AI on joint-position data teaches it to produce movements that are geometrically plausible. The skeleton visits the right coordinates at roughly the right times. But it doesn't teach the AI anything about how that motion was produced — what the mover intended, how the effort was organised, whether the movement was alive with internal motivation or mechanical in its execution.

This is why AI-generated movement has historically looked slightly wrong even when it's technically correct. The joints go where they're supposed to go. But something is missing in the in-between — the weight, the intention, the quality that makes a movement feel inhabited rather than performed by a puppet.

LaMoGen: Using Laban's Symbols as a Bridge

Researchers at the intersection of AI and movement science are now using Laban notation as an explicit intermediate representation — a bridge between natural language and movement quality.

The most direct example is LaMoGen (arXiv:2603.11605, March 2026), which introduces a framework called "LabanLite": a streamlined version of Labanotation in which each atomic body-part action is encoded as a discrete Laban symbol paired with a textual template. The system works in three steps:

  1. A large language model reads a natural language description of a desired movement and generates a LabanLite sequence — a symbolic representation of what the movement should feel and look like in Laban terms.

  2. The LabanLite sequence is decomposed into body-part instructions, each linked to a specific Laban symbol and its associated effort and shape qualities.

  3. A motion generation model converts the LabanLite instructions into actual joint-position sequences, guided by the qualitative constraints the symbols encode.

The crucial innovation is step 1: making the intermediate representation interpretable. Instead of asking an AI to jump directly from "perform a sorrowful, heavy descent" to joint coordinates, LaMoGen asks the language model to first translate that instruction into Laban symbols — a language that preserves the qualitative content of the description — and then generate motion from those symbols.

This matters because the Laban representation is interpretable to humans. A choreographer looking at the LabanLite output can check whether the AI's "translation" of the instruction is correct before the motion is generated. They can edit it. They can substitute different Laban symbols to modify the effort quality while keeping the spatial shape intact. The symbolic intermediate layer makes the generation process legible and adjustable, rather than a black box.

What This Means for Somatic Practice

For somatic practitioners, there is something philosophically significant about this development — not just technically interesting.

Somatic training is, in part, the cultivation of exactly the discriminative capacity that Laban notation attempts to encode. A practitioner who has trained in Body-Mind Centering, the Alexander Technique, Feldenkrais, or Laban Movement Analysis itself develops the ability to perceive and modulate effort quality, to notice the difference between free and bound flow in a student's movement, to adjust their own weight and timing to shift the quality of a phrase. This is the knowledge that takes years to develop and that cannot be read off from a video recording.

Laban notation was Laban's attempt to make that knowledge transmissible — to give teachers, students, and researchers a shared vocabulary for talking about movement quality in precise terms. The notation has been used in this way for a century, with varying degrees of adoption.

What the LaMoGen approach suggests is that AI systems can now participate in this vocabulary — not yet as fluent speakers, but as capable readers and writers of basic Laban symbols. A system that generates movement guided by Laban effort qualities is, in a modest but genuine sense, a system that can be instructed in somatic terms.

The remaining gap is the one identified in previous work on predictive architectures and proprioceptive sensing: Laban notation describes effort quality from the outside (the observer's perception of the movement's quality), while somatic practice cultivates effort quality from the inside (the mover's felt experience of their own movement). An AI system trained on Laban annotations is learning the observer's vocabulary. To learn the mover's vocabulary, it would need to be trained on the felt signals — the EMG pre-activation, the weight distribution, the joint tension — that give rise to those observable qualities.

Laban gave AI a language for movement quality. The remaining question is whether AI can learn to speak it from the body's own perspective, rather than from a seat in the audience.

Further Reading

Laban, R., & Lawrence, F. C. (1947). Effort: Economy in body movement. MacDonald & Evans.

Newlove, J., & Dalby, J. (2004). Laban for actors and dancers. Nick Hern Books.

Zhao, M., et al. (2026). LaMoGen: Language to motion generation through LLM-guided symbolic inference. arXiv:2603.11605. https://arxiv.org/abs/2603.11605