When AI Learns the Laws of Motion — and What That Still Doesn't Tell It About the Body
A quiet revolution is making AI movement more physically real. But physical realism and somatic authenticity are not the same thing.
Watch a video of a person tripping on a step. Without thinking, you know what is about to happen: the weight will shift forward, the arms will come up, one foot will scramble to catch. You feel something in your own body — a small lurch of anticipation — before the person even starts to fall.
This felt anticipation is not magic. It is physics, encoded in decades of your own experience of bodies moving through a gravitational world. You know what happens to mass when momentum overtakes balance because you have been that mass. You carry the laws of motion in your body.
For a long time, AI motion generation systems did not. They learned to produce movement that looked right from the outside — visually plausible, statistically consistent with the movements humans perform — but that had no internal physical logic. If you asked such a system to generate a person tripping, it would produce something that resembled a trip. But the weight distribution would often be wrong, the recovery timing improbable, the arm positions physically impossible given the forces involved. The more technically demanding the movement, the more obvious the gap.
This is changing. And understanding how — and why it is only a partial solution — matters for anyone interested in what AI can and cannot do with movement.
The Physics Turn in Motion Generation
In the past year, a wave of research has given AI motion systems an explicit understanding of physics. Rather than training only on recorded human movement, these systems now train on physics simulations — environments where forces, friction, gravity, and momentum are modelled mathematically and verified at every timestep.
One approach, called PhyCo, trains a video generation model on hundreds of thousands of simulated clips in which physical parameters are systematically varied: the friction of a surface, the elasticity of a collision, the force applied to an object. The resulting model can generate video of movement that not only looks plausible but obeys the actual physics of the specified environment. You can tell it "this surface has low friction" and the generated person will slip; tell it "high friction" and they will push off with appropriate grip.
Another approach, InterPhys, focuses specifically on how bodies interact with their environments at contact points — where hand meets table, where foot meets floor, where two dancers' bodies connect. Rather than treating contact as a positional constraint ("the hand should be near the surface"), InterPhys models contact as a force exchange ("these are the forces acting at the point where hand meets surface"). The difference matters enormously for the realism of the result: a model that knows about contact forces will generate a hand that pushes convincingly, not just arrives at the right location.
A third system, PhysiGen, addresses one of the most persistent artefacts in AI-generated movement involving multiple people: body interpenetration — the ghostly phenomenon where one animated body appears to pass through another. By explicitly modelling the collision constraints between bodies, PhysiGen dramatically reduces this effect, making two-person movement much more physically credible.
Together, these represent a genuine leap. Motion generation is becoming physically grounded in a way that was not possible two years ago.
What Physics Explains — and What It Does Not
Here is the honest accounting of what this progress achieves and what it does not.
Physics explains the constraints on how a body can move. A person cannot accelerate faster than their leg muscles can produce force. A person in a low-friction environment will slide differently than in a high-friction one. A body in contact with another body exchanges forces that both parties must respond to. These are hard constraints, and generating movement that respects them makes AI output significantly more believable.
What physics does not explain is why a particular movement happened at a particular moment, with the particular quality it had. A somatic practitioner making a reaching gesture is not primarily solving a physics problem. They are following an impulse, responding to a sensation, extending toward something with an intention that preceded the movement in their nervous system. The physics is the medium through which that intention expresses itself — but it is not the intention.
This is not a subtle distinction. In body-based practices — from Contact Improvisation to martial arts to somatic therapy work — the felt quality of a movement is the primary content, not its external shape or its physical correctness. A reaching gesture done with longing feels different from the same gesture done with caution, even if the joint angles and forces are almost identical. Observers can perceive this difference reliably. Movement practitioners are trained specifically to sense it.
Current physics-based AI systems can generate movement that is physically correct. They cannot generate movement that is motivated — in the sense that practitioners use that word. The motivation (the intention, the sensation, the emotional-kinaesthetic content) is precisely what somatic practices are about, and it is precisely what neither training data nor physics simulation contains.
A Concrete Example: The Stumble
Return to the tripping person. A physics-aware AI can now generate a very good stumble: the weight shifts correctly, the forces at the foot are consistent with the surface, the recovery arc follows plausible biomechanics.
But there are many qualitatively different stumbles. The stumble of someone who is distracted and suddenly snaps into alertness. The stumble of someone who is exhausted and doesn't quite have the reflexes to recover cleanly. The stumble of a dancer who uses the trip as an intentional departure from the choreography into improvisation. Each has a different physical signature — different latencies in the recovery, different quality of the arm response, different tension in the torso — and these differences arise from the internal state of the mover, not from the physics of the surface.
A somatic practitioner working with movement quality can train themselves to produce different stumbles from the same physical input. They are not overriding the physics; they are populating the physics with intention. This is what somatic intelligence is: the capacity to give felt, intentional content to movements that physics only constrains.
AI systems can now follow the constraint. The intention is still the mover's to contribute.
APA Further Reading
Narayanan, S., Jiang, Z., Narasimhan, S., & Chandraker, M. (2026). PhyCo: Learning controllable physical priors for generative motion. arXiv:2604.28169. https://arxiv.org/abs/2604.28169
Sheets-Johnstone, M. (2011). The primacy of movement (2nd ed.). John Benjamins. https://doi.org/10.1075/aicr.82
Tversky, B. (2019). Mind in motion: How action shapes thought. Basic Books.
Xing, C., Mao, W., & Liu, M. (2026). InterPhys: Physics-aware human motion synthesis in a dynamic scene. arXiv:2605.01036. https://arxiv.org/abs/2605.01036