Reinforcement learning in the studio – what dance teachers already do
AI reinforcement learning formalizes what dance classrooms do physically: try, feedback, repeat. Thorndike first, Sutton and Barto later – and back into practice.
Last week, adagio tempo, quiet piano: relevé at the barre, heel wavering on the hold. No biomechanics lecture. Two fingers on the shin, light outward pressure, three slow repetitions – on the fourth she held it alone without my hand. No handout.
That is most of my teaching. Demonstrate, touch, correct, repeat. Paper explains what is intended. The body learns through attempts.
AI did not invent this
Reinforcement learning in AI grows out of learning psychology. Sutton and Barto write in Chapter 1.6 of their RL introduction that the trial-and-error thread starts with Thorndike’s Law of Effect (1911): actions followed by satisfying outcomes become more likely (Thorndike, Animal Intelligence). Instrumental learning – behavior changes the situation, reward shapes the policy – was underrepresented in engineering until Klopf, Sutton, and Barto formalized it.
RL is not a metaphor we retro-fit onto children. It is a formalization of what organisms already do.
Motor learning: why talk alone fails
Dance research and motor learning (Schmidt & Lee, Motor Control and Learning; IADMS resources) distinguish intrinsic feedback (balance, tension) from augmented feedback from the teacher.
Two forms matter:
- Knowledge of results (KR): “You fell to the right.”
- Knowledge of performance (KP): “Your right hip lifted, so you lost your axis.”
Systematic reviews show KP is not always superior – context matters (skill level, task). For complex movement like dance, KP plus KR is often needed, and for novices prescriptive KP (“turn out the heel earlier”) often beats outcome-only description. The body needs information about movement quality, not just failure.
Demonstration, tactile input, and repetition are not a relic. They are how the body actually receives information in class.
The studio loop (RL vocabulary, human meaning)
| RL term | In class |
|---|---|
| State | Posture, fatigue, music, floor |
| Action | The attempted step |
| Reward | Stability, “yes, like that,” applause, internal confidence |
| Policy | Your default entry into a pirouette |
| Exploration | Improv, testing a different quality |
| Exploitation | Fixed version before an exam |
The reward is rarely a number. It is tone, pause, sometimes silence. Still, the body optimizes through repetition.
Reapplying four RL ideas to humans
1. Reward shaping (intermediate reinforcement)
In pirouettes I do not reward only “you stayed up.” I mark turnout before push-off, closed arms, early eyes.
Credit assignment: Without those intermediate markers, the body does not know which link in the chain caused success.
2. Exploration vs. exploitation
Rehearsal: variation and improv. Performance week: a stable policy. Many students exploit too early – drilling a half-finished version until it freezes.
3. Delayed reward
Premieres and exams give sparse feedback. Class needs dense rewards: small corrections per repetition, not one sentence at the end of the hour.
4. Model-based vs. model-free
Sometimes you need imagery and anatomy (planning). Sometimes pure drill until it runs automatically. Effective teachers switch between both; staying in one mode slows progress.
Second example: center adagio
A student falls backward in développé, standing knee buckling slightly. KR would be: “You lost balance.” I say: “Your standing knee is giving way – feel the thigh rotate outward.” Then five repetitions without new words, only a nod or brief pause when it lands. Terminal feedback plus repetition – the loop motor-learning literature describes for skill acquisition.
What I am not claiming
I am not claiming students are GPT models or that PPO runs in the brain. The mapping table is pedagogical language to see where we intervene more clearly.
RL has also fed back into psychology (temporal-difference models of conditioning). The direction is not one-way.
Further reading
- Sutton & Barto: History of RL (1.6)
- Thorndike: Law of Effect (Animal Intelligence)
- IADMS: Motor Learning and Teaching Dance (PDF)
- PMC: Augmented feedback systematic review
Takeaway
Dance is too physical for theory alone. RL in AI is too abstract without a body. The middle is: try – feel – adjust – repeat. Thorndike described it. Sutton and Barto formalized it. You do it at the barre every day.
I can follow with a pirouette class designed as explicit reward shaping if readers want that next.