Amy LaViers: Teaching Robots to Mean What They Do
The director of the Robotics, Automation, and Dance Lab on Laban notation, expressive machines, and why somatic training is a prerequisite for building AI that moves well
Amy LaViers has spent the past fifteen years asking a question that most roboticists have never considered: what does a robot's movement mean? Not how does it function — that is a solved problem for most locomotion tasks — but what does the quality of how it moves communicate to a human observer, and how can that quality be deliberately designed?
Her route to this question runs through dance. LaViers trained in dance before moving into engineering, and when she arrived in robotics, she carried with her something the field largely lacked: a vocabulary for movement quality. That vocabulary was Laban Movement Analysis — the framework developed by Rudolf Laban in the early twentieth century to describe the effort, shape, and spatial qualities of human movement. Laban had built the vocabulary for dancers and movement educators. LaViers has spent her career extending it to machines.
Your lab is called the Robotics, Automation, and Dance Lab. The RAD Lab. How do you explain that combination to someone who hasn't thought about it before?
Most people think of robotics and dance as completely separate domains. Robotics is engineering: you build something that accomplishes a task, and the measure of success is whether the task gets done. Dance is art: you make something that communicates, and the measure of success is whether the communication lands. What I'm interested in is the cases where those two things aren't separate. Where how the robot moves is part of what it accomplishes. Where the quality of the movement is the task.
Think about a robot that assists an elderly person with physical rehabilitation. The mechanical outcome is: the arm completes the prescribed range of motion. But there's a huge difference between a robot that guides that motion in a light, responsive, attentive way and one that mechanically forces the limb through the range. The patient can feel the difference. The relationship between patient and robot — the quality of that interaction — depends on movement qualities that standard robotics metrics don't even try to measure.
Dance is a centuries-old discipline for developing sensitivity to exactly those qualities. My work is about bringing that sensitivity into engineering.
How does Laban Movement Analysis function as a practical tool in your work?
Laban analysis gives us a shared vocabulary for talking about movement qualities in precise, transmissible terms. The Effort dimension — the four motion factors of Weight, Space, Time, and Flow — provides eight fundamental quality contrasts: strong vs. light, direct vs. indirect, sudden vs. sustained, bound vs. free. These are not aesthetic judgments; they are descriptions of observable, measurable properties of movement. And they map onto felt experience in ways that other movement descriptions don't.
When I say a movement is "bound," I mean that the mover is actively managing the flow of movement energy — there's a sense of held containment, preparedness, potential. When I say it's "free," I mean the movement energy is released and ongoing, self-propagating. Those are distinct felt qualities. A dancer can reliably distinguish them and deliberately produce them. A robot, properly designed, can also produce them — and when it does, humans reliably perceive the difference.
The practical tool is notation: we can write down a movement in Laban terms and use that notation as a specification for a robot's movement quality, not just its path. That's what conventional trajectory planning can't do. It can specify position and velocity. It can't specify whether the movement should feel sudden or sustained, or whether the approach should be direct or circuitous. Laban notation can.
Your book with Catherine Maguire, Making Meaning with Machines, came out through MIT Press. What was the core argument you were making there?
The book is a practical argument. We're saying: if you want to design machines that communicate meaningfully — that move in ways that are legible, appropriate, and expressively intentional — you need to develop the somatic and analytical capacities to perceive and describe movement quality, not just movement geometry.
Most engineers learn to describe movement in terms of coordinates, forces, and time. Those descriptions are necessary but not sufficient for expressive design. What they leave out is the felt dimension — the dimension that Laban analysis was built to capture. Our book is a primer in that dimension, written specifically for designers and engineers who haven't trained in dance or somatics.
We're also making a claim about what somatic training does. It's not just about acquiring a vocabulary. It's about developing a perceptual capacity — the ability to notice qualities of movement that you would otherwise not consciously register. That capacity is hard to acquire from a textbook. It requires embodied practice: moving, receiving feedback, developing sensitivity through your own body's experience.
We try to make that process tractable for people who haven't trained in dance. But we're honest about the fact that it takes time and genuine physical engagement. You can't shortcut it by reading more carefully.
Where does AI fit into your current work?
It's increasingly central. The most interesting development for me is that AI motion generation is now capable enough that the question of quality is urgent rather than theoretical. We can generate movement. The question is whether we can generate movement that means something.
What I find hopeful is the work being done on using Laban notation as an intermediate representation in text-to-motion systems. If you can specify a desired movement quality in Laban terms and have a model generate motion that satisfies those specifications, you've created a channel from qualitative intention to physical output. That's what expressive design needs.
What I find concerning is how easy it is to generate movement that looks expressive without being expressive in any principled sense. A system that has learned statistical patterns in human movement can produce outputs that superficially resemble intention without actually encoding it. The difference matters — not as an aesthetic judgment, but as a functional one. A rehabilitation robot that looks attentive but isn't calibrated to be actually responsive is a deception. That's not a good outcome.
What would it take for AI motion systems to generate movement that is genuinely expressive in the sense you mean — not just statistically likely to appear expressive?
It requires grounding the training in the signals that encode quality, not just the signals that encode appearance. Effort quality is produced by the body's neuromuscular organisation before it becomes visible in the body's shape. If you train a model only on video — on the appearance of the resulting movement — you're working at the downstream end of that production process. The upstream signals that shaped the output aren't there.
This is why I'm interested in what wearable sensing can do. EMG, inertial measurement, proprioceptive data — these are signals that are closer to the neuromuscular production process. They carry information about effort quality that has been dissipated by the time the movement is visible to a camera.
The Laban effort factors are ultimately descriptions of neuromuscular organisation: bound flow is high co-contraction; free flow is low antagonist activity; Weight relates to force production; Time relates to the rate of change of force. If you train a model on the signals that directly encode those neuromuscular states, you have a chance of building a system that generates movement from the inside out — from the quality intention to the visible expression — rather than from the outside in.
You mentioned the somatic training required to perceive movement quality. Is there a role for AI in developing that perceptual capacity, rather than just expressing it?
That's a question I think about a lot. There are some clear applications: systems that provide feedback on movement quality during practice, systems that generate reference movements for a given quality specification so that a learner can compare their own movement against a target. Those are useful pedagogical tools.
Where I'm more cautious is the idea that AI can substitute for the embodied development of perceptual capacity. The reason somatic training works is that it changes the body — it changes the nervous system's sensitivity to proprioceptive and kinaesthetic signals. That change happens through physical experience. An AI that provides analysis from the outside doesn't change the body. It might accelerate certain aspects of learning — giving more granular feedback more frequently — but it can't replace the accumulation of embodied experience that somatic training provides.
My hope for AI in this space is as a collaborator and amplifier for practitioners who already have that training, not as a substitute for it. The practitioner provides the embodied intelligence; the AI extends its reach — generating, responding, notating, archiving. That's a genuinely interesting partnership. What I want to avoid is the assumption that the AI can take over the somatic intelligence entirely and that the practitioner just provides a prompt.
Amy LaViers is Director of the Robotics, Automation, and Dance (RAD) Lab in Philadelphia. She is co-author (with Catherine Maguire) of Making Meaning with Machines: Somatic Strategies, Choreographic Technologies, and Notational Abstractions through a Laban/Bartenieff Lens (MIT Press, 2023, open access). Her research is published in the Journal of Laban Bartenieff Movement Studies, IEEE Transactions on Human-Machine Systems, and the Annual Review of Control, Robotics, and Autonomous Systems.
References
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/
LaViers, A. (2025). Robots and dance: A promising young alchemy. Annual Review of Control, Robotics, and Autonomous Systems, 8, 323–350. https://doi.org/10.1146/annurev-control-060923-100542