Using Laban Effort Vocabulary to Prompt AI Motion Generation
A practical guide for somatic practitioners who want to give AI systems qualitative instructions, not just geometric ones
Who This Is For
This guide is for movement practitioners — dancers, somatic educators, physical theatre makers, bodyworkers — who are beginning to use AI motion generation tools in their practice or research. It assumes familiarity with Laban Movement Analysis, particularly the Effort dimension. It does not assume any programming knowledge.
The central problem this guide addresses: most AI motion generation systems accept instructions in terms of actions ("walk," "reach," "turn") or text descriptions ("a person moving slowly and heavily across the room"). They do not natively speak the language of Laban effort — the precise qualitative vocabulary that somatic practitioners use to distinguish a bound, sudden reach from a free, sustained one.
This guide explains how to translate between the two: how to express qualitative intentions in forms that current AI tools can receive, and how to interpret what you get back.
Part 1: The Vocabulary Gap
When you ask an AI motion generation system for "a heavy, sustained descent," it will probably generate something that looks heavier and more sustained than its default output. The system has learned statistical associations between these words and the movement patterns that humans, when annotating motion data, tend to label with them.
What the system has not done is compute the effort quality from first principles — generate the appropriate neuromuscular organisation and then produce the visible result from it. It has learned that certain joint-position patterns co-occur with certain words. This is a meaningful but limited correspondence.
The practical implication: your qualitative language will influence the output, but the influence is indirect and statistical rather than precise and principled. The word "bound" shifts the distribution of generated motion toward patterns that annotators labelled "bound." It does not guarantee that the generated motion has the structural features — the high co-contraction, the preparedness, the held potential — that constitute bound flow in somatic terms.
This is not a reason to stop using qualitative language with AI tools. It is a reason to approach the results as dialogue rather than execution: you provide a quality intention; the system offers an interpretation; you respond to that interpretation with another quality specification. The output is a proposal, not a transcript.
Part 2: Translating Effort Qualities into Prompts
The Laban Effort dimension has four motion factors, each with two poles. Here is how each translates into language that current AI tools are more likely to respond to:
Weight: Strong ↔ Light
Strong — Synonyms that work in text prompts: heavy, grounded, forceful, weighty, pressing, powerful, deliberate. Phrase patterns: "moving as though pressing against resistance," "landing with full weight," "each step grounded and deliberate."
Light — Synonyms: delicate, buoyant, feathery, floating, airy, gentle, barely touching. Phrase patterns: "barely touching the ground," "as though the body were weightless," "moving with the lightest possible contact."
Practitioner note: Weight effort is often the easiest to get AI systems to respond to because it has strong correlates in visible kinematics (speed of descent, impact force). Strong/light distinctions are usually the most reliably rendered.
Time: Sudden ↔ Sustained
Sudden — quick, sharp, instantaneous, burst, snap, flick, immediate. Phrase patterns: "a sharp sudden change of direction," "movement that arrives all at once," "as though responding to an unexpected stimulus."
Sustained — slow, continuous, unbroken, flowing through time, unhurried, drawn out. Phrase patterns: "a sustained arc through space," "movement that never quite arrives," "the gesture unfolds over a long, unbroken phrase."
Practitioner note: Temporal qualities are well-represented in motion data (directly measurable as velocity profiles), so sudden/sustained instructions tend to be well-followed. Be specific about what is sudden or sustained — the whole phrase, a specific body part, the initiation, the completion.
Space: Direct ↔ Indirect
Direct — straight, linear, targeted, single-focus, moving straight toward, without detour. Phrase patterns: "moving in a straight line toward the destination," "single-focus, no deviation from the path."
Indirect — curving, circuitous, multi-focus, weaving, meandering, exploratory. Phrase patterns: "the arm traces a winding path through space," "moving without a fixed destination, attending to the whole spatial environment."
Practitioner note: Space qualities are harder for AI systems to respond to than Weight and Time, because they are less directly encoded in standard joint-position data. Using spatial imagery (curving, tracing, spiralling) often works better than the Laban terms directly.
Flow: Free ↔ Bound
Free — uninterrupted, released, ongoing, self-propelling, fluid, unstoppable. Phrase patterns: "the movement continues of its own momentum," "released and flowing, hard to stop," "energy moving freely through the whole body."
Bound — controlled, held, careful, contained, reversible, ready to stop at any moment. Phrase patterns: "movement held in check," "each gesture contained and controllable," "moving as though through thick air, every moment reversible."
Practitioner note: Flow is the hardest Effort quality for AI systems to render precisely — the distinction between free and bound is most visible in micro-temporal texture (acceleration profiles, micro-hesitations) which is often lost in standard motion capture data. Use kinetic imagery (momentum, resistance, containment) rather than the words "free" or "bound" directly.
Part 3: A Practical Workflow for Quality-Led Generation
Step 1: Define your Effort intention in Laban terms
Before writing a prompt, identify the specific Effort qualities you want. Use the four motion factors: Weight, Time, Space, Flow. Not all four need to be specified — sometimes a single quality is the expressive focus.
Example: You want a phrase that is light, sustained, and indirect — something hovering, wandering, unhurried.
Step 2: Generate a natural-language translation
Translate your Effort specification into the richest natural-language description you can construct. Use imagery, physical analogy, contextual scenario.
Example prompt: "A figure moving as though suspended slightly above the ground, buoyant and barely touching, tracing a long, wandering arc through space without arriving at any fixed point. The movement is unhurried, almost dreamlike, each transition flowing into the next without interruption. The body attends to the whole environment simultaneously rather than moving toward a single destination."
Step 3: Generate and analyse the output in Effort terms
When you receive generated motion — whether as video, skeleton animation, or a live system output — analyse it using the same Effort framework you used to prompt it. Ask: does this output have the Weight quality I intended? The Time quality? The Space quality? The Flow quality?
Be specific. Often an AI will render some Effort qualities accurately and miss others. Identifying which is which tells you exactly what to adjust in the next prompt iteration.
Step 4: Iterate toward the intended quality
Your second prompt is a correction, not a repetition. Be specific about which quality was not rendered as intended.
Example correction: "The weight and timing are right — keep those. But the path through space is too direct; the movement keeps resolving into a straight line toward a point. Make the spatial quality more genuinely indirect: curving, multi-focus, without arrival."
Step 5: Use the output as a starting point, not an endpoint
The generated motion is a proposal. In somatic practice, the most useful outcome is often not a finished phrase but a movement idea you can inhabit, modify, and make your own. Use the AI output as a score suggestion — a kinetic proposal that your body can respond to, resist, and extend.
Part 4: Working with Currently Available Tools
Text-to-motion systems (e.g., HY-Motion 1.0, MDM-based models) Accept text prompts and generate skeleton animations. The workflow above applies directly. Best for: exploring quality variations on a basic action, generating movement material for choreographic score development.
Reference-video-driven systems (e.g., Kling 2.6 Motion Control) Take a reference movement video and transfer its qualities onto a new character or context. Use as a quality-transfer tool: record yourself performing a phrase with specific Effort qualities, then generate how those qualities look on a different body or in a different visual context.
In-betweening tools (e.g., systems based on arXiv:2605.12778) Generate the movement between two keyframe poses. Specify your keyframes (starting and ending positions) and add quality language to describe the path. The quality language guides how the transition unfolds, not just where it goes.
What to Expect: An Honest Assessment
Current AI tools respond to Effort qualities unevenly. Weight and Time tend to be rendered more reliably than Space and Flow. Complex Effort combinations — a phrase that is simultaneously light, sudden, and bound — are harder to achieve than single qualities.
The gap between a Laban-informed quality intention and what AI systems can currently produce reflects the data gap: most systems are trained on data that records movement appearance without Effort annotation. As Laban-annotated datasets grow and as researchers like Amy LaViers develop tools for labelling Effort qualities at scale, this gap will narrow.
For now, the most useful stance is one somatic practitioners already know: work with what arises. The AI's interpretation of your quality intention is a collaborator's response, not a failure to follow instructions. Stay curious about the gap between what you intended and what emerged. Sometimes the most interesting material is in the discrepancy.
Further Reading
LaViers, A., & Maguire, C. (2023). Making meaning with machines: Somatic strategies, choreographic technologies, and notational abstractions through a Laban/Bartenieff lens. MIT Press. https://mitpress.mit.edu/9780262546126/making-meaning-with-machines/ (open access)
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