PAI-230 · Education

Physics-Informed Physical AI

The physical world will not be solved by data alone: action-labeled data is scarce, and a black-box net that has never heard of energy or momentum extrapolates dangerously. This course teaches the alternative, inject the physics you already know into a learner, at one of three places: the data (guided), the loss (informed), or the architecture (encoded). You build every method on-device in the browser, driving the real ferromotion Rust library compiled to WebAssembly: reverse-mode autodiff, physics-informed neural nets, Lagrangian and Hamiltonian nets, Neural ODEs, structure-preserving integrators, equation discovery, and differentiable control. Physics → AI → Physical AI, one runnable rung at a time.

Skilled → Frontier·10 modules · 16 live labs·24 lessons
▶ Start the course ← All courses
THE HORIZON

Where this sits, and what moves it.

Binding constraint · Action-labelled samples per unit of dynamics learned. Physics you already know -- a conservation law, an equation of motion -- is information you do not have to pay for in data, and the exchange rate between the two is what this whole field is trading on.

Was impossible

A learner either knew physics because someone hard-coded it or learned everything from scratch. There was no principled place to put a conservation law you already believed, so it was either an assumption or an absence.

Is probable

There are three places -- the data, the loss, the architecture -- and this course builds all three plus discovery. the ordering matters: a loss penalty is soft, so a low loss is not a correct answer, which is why the verify-first module exists rather than being an appendix.

Becomes possible

The frontier is architectural rather than penalty-based: structures that cannot violate the physics regardless of what the optimiser does, so the guarantee survives training. Where that is hard is exactly where the lumped model breaks -- fluids, contact, deformation -- and that is where the course spends its last modules on purpose.

Every hard thing was impossible until the constraint that made it impossible was named. How we read a frontier →

Module 6

Physical AI: closing the loop

Put the models to work. A model you can differentiate is a model you can control by gradient descent, so tune a controller by backpropagating through the loop, then close the whole arc: learn a model from data and train a controller inside it.

  1. L3Control by gradientYou have spent the course building models you can differentiate. What does that buy when you close the loop?Tune a controller by backpropagating a trajectory cost through the closed loop, the payoff of a differentiable model. See why one gradient per rollout beats reinforcement learning's thousands of episodes.
  2. L3Learn in imaginationYou are handed an unknown plant. You identify a model from data, tune a controller entirely inside that model, then deploy. Where does the risk sit?Close the whole loop: identify a model of an unknown plant from data, tune a controller entirely inside that learned model, and deploy it on the real system. The complete physics-informed pipeline, on-device.
  3. L3The industry agrees: structure-as-differentiation, at scaleA world model predicts the next video frame directly, pixel by pixel, from the current frame and an action, no explicit position, velocity, or mass anywhere inside it. Can it violate conservation of momentum during a predicted collision?See that the principle this course has built lesson by lesson (a physics-structured, differentiable model, identified and controlled by gradient) is the same bet three 2025–2026 efforts are placing at industrial and research scale: Newton (GPU differentiable physics, contributed to the Linux Foundation by Disney Research, Google DeepMind, and NVIDIA), gradient-based system identification on real robots using MJX, and OrbiSim's argument that a world model should itself be a differentiable physics engine rather than a generative video predictor.
Module 8

The model meets reality

Close the loop with the physical world: calibrate the model from recorded data by exact gradients through the dynamics, and find out what the numbers you grade a grasp with leave out.

  1. L3The model meets realityEvery model so far was right by construction: you knew the equations and the parameters. A real robot hands you neither. What do you do?Calibrate a robot's masses and friction from recorded torques by exact gradients through the dynamics, real-to-sim, the single most validated use of differentiable physics.
  2. L4The Grasp Number Lies TwiceYou certify a grasp with a force-closure quality and a stick-or-slip flag. Both come back fine. What have you actually established?Take the two numbers everyone grades a grasp with - a force-closure quality and a stick-or-slip flag - and find what each one hides. You will discover that a coplanar grasp is full rank in six dimensions rather than degenerate; that the force-closure number is a direction-sampled UPPER bound whose value falls threefold as you probe more finely, so a claim read off it can be a property of the sampler rather than of the grasp; and that a contact reported as stuck has usually already lost a quarter of its stuck area. Pass = you read the rank, read how far the quality metric still moves under finer sampling, and find the load at which a stick-area monitor warns against the load at which a Coulomb monitor does.
  3. L3The Robot Is the FileSomeone hands you a robot description. It parses, the articulation moves, every number is plausible, nothing warns. Where is the bug?Price three conventions in a robot description in newton-metres, and find out that the pose you test in decides which of them you can see. You will read a joint limit that means something different depending on the joint type, watch a length-scale error change a torque by a factor that is not the length ratio, flip a header token and move gravity, and discover a pose where a hundred-fold error is exactly invisible. Pass = you state all three errors as numbers and explain why one of them cannot be undone by rescaling the answer.
Module 10

Verify-first: don't trust, certify

A physics-informed net only SOFTLY penalizes the physics, so a low loss is not a correct answer. Turn a trained solution, and a trained policy, into ones you can trust by computing a certificate on the output: the Institute's verify-first spine, from a PINN's residual bound to an energy that certifies every action.

  1. L4Don't trust a PINN, certify itA physics-informed net's training loss went to nearly zero. Does that guarantee its solution is correct?A PINN drives a residual toward zero, a soft penalty, not a guarantee. Compute an a-posteriori certificate: a rigorous bound on the true error from the trained net's own residual, needing no true solution. Then watch the same idea become an energy certificate on a controller and on a frozen policy. Soft → structural → proven.
  2. L4The Gradient That LiesYou take a contact you can solve exactly and refine the simulator's timestep. What happens to the gradient it returns?Measure, on a contact you can solve exactly, that the gradient a differentiable simulator returns does not converge as the contact stiffens - it diverges as the square root of stiffness and points the wrong way over most of the usable range. Then find out WHICH PART of the simulator is responsible, because the obvious answer is the wrong one. Pass = you read the divergence exponent off the lab, find a stiffness where the sign is wrong, and show the same contact model getting the right answer once the integrator changes.
  3. L4The Reference That Shares Its MachineryYou want to split the contact-gradient error into its causes. Your reference solver shares the same integrator. What does the comparison measure?Take a measurement you already trust and find out what it actually measures. You will split the contact-gradient error into the part the integrator causes and the part the contact model causes, discover they move in opposite directions, and learn the two ways a reference lies to you: by sharing machinery with the thing it scores, and by having a noise floor above the number you wanted to quote. Pass = you name the integrator share and the model share at 1e6, and say why an answer that moves when you change the probe is not an answer.
  4. L4What a Certificate Cannot ClaimA verification returns UNDECIDED. What does that tell you?Learn to read a verification result as a claim with preconditions rather than a number. You will make the same system return Certified, Refuted and Undecided; discover that one of those transitions involves no change to the geometry at all; find a green verdict that means nothing; and find a fourth failure the lab uncovered in itself, a bound that was only valid in a region nobody checked. Pass = you produce all three verdicts, explain which control caused each, and say what the precondition check found.