The Two Operating Systems: What a Body Carries That No Dataset Contains
Companies and academic centres are converging on the view that a sufficient volume of recorded experience will unlock Physical AI, and the strongest evidence is real: a log-linear scaling law from twenty thousand hours of egocentric human video. This course tests the thesis mechanically, on the Institute's own bench. You will unplug a motor and measure the stiffness at which a body balances itself, show that two demonstrations on that body certify what a hundred and twenty-eight on a stiff one cannot, build three perfect local models that provably admit no joint model and show that more data cannot create one, measure a grammar in which seeing every rule takes ten samples and seeing every sentence takes more than sixty thousand, break a simulated fly by inverting the sign of its own efference copy, and train the same clone twice from the same demonstrations, once from the actions the demonstrator logged and once from actions an observer reconstructed from the motion, and watch the second collapse from a perfect score to zero the moment the body holds a steady load. Every number in every lesson is printed by the lesson's own cells, and the research bench behind them registered its predictions before running.
▶ Start the course ← All coursesWhere this sits, and what moves it.
Binding constraint · Observability. The substrate does its work below the first transducer: torque exchanged directly between a compliant body and its load never crosses a sensor, so no volume of recorded observation contains it, and an observer reconstructing actions from motion books that work to the wrong author. The deciding quantity is the share of the loop's output that originates below the sensors, and on this course's bench that share moves the outcome more than a sixty-four-fold change in data volume.
The demonstrations existed for decades and were read as curiosities: a decerebrate cat walks on a moving treadmill, and a passive walker with no motor and no controller walks down a slope at a fraction of a powered humanoid's cost of transport. What changed is that compliance became something you could purchase: series-elastic actuators, variable-stiffness joints and back-drivable drives turned the substrate from an anatomical observation into an engineering budget line.
The data thesis holds a real scaling law and feeds production robot foundation models, and it is measurably incomplete. On the bench behind this course, two demonstrations on a compliant body certify the whole operating region while five hundred and twelve on a stiff one reach a quarter of it, and action labels reconstructed by an observer go from benign to fatal the moment the body holds a steady load. The structural and statistical legs are published proofs; the embodied leg is measured at 256 seeds with every prediction registered before its run.
Watch for the loop being logged rather than filmed. A major egocentric or teleoperation dataset shipping per-frame motor commands and a declared substrate model alongside its video, a scaling paper reporting operating region separately from tracking error, or a policy architecture that consumes an efference copy as a first-class input: any one of these is this course's thesis arriving in production, and all three exist today as parts.
Every hard thing was impossible until the constraint that made it impossible was named. How we read a frontier →
The body underneath
Measure what the substrate carries before any policy exists: a body that balances with its motor unplugged, a demonstrator whose competence is mostly the body's, and a clone for which two demonstrations on the right body outbuy sixty-four times the data.
- L2Balance with the motor unpluggedThe motor is unplugged and the pole starts 0.05 radians off vertical. What keeps it up, if anything?Show that a pivot spring above a threshold stiffness balances the pole with zero control, and locate the threshold two ways: by simulation and by formula.→
- L3The demonstrator's secretThe identical expert controller drives four bodies from free pivot to very stiff. Which body gives it the best tracking of the moving target?Measure how much of an expert demonstrator's competence belongs to its body, by running the identical controller on four bodies and reading region and tracking separately.→
- L3Two demonstrations against one hundred and twenty-eightAt kp=0 the clone gets sixty-four times more demonstrations, 2 to 128 episodes. How much operating region does that buy?Clone the expert from camera observations at two data volumes on two bodies, and measure which axis moves the outcome: data or substrate.→
Three mathematics, one claim
Establish the structural and statistical legs: coherence across local models is a gluing condition no data volume moves, token-level learning pays exponentially where predicting your own latents pays a constant, and biology named the mechanism in 1950.
- L3The model that cannot existThree pairwise models, each perfect: A-B at +0.9, B-C at +0.9, A-C at minus 0.9. Does a joint model of all three exist?Construct three perfect local models that admit no joint model, and show the obstruction is a spectral threshold that more data per pair cannot cross.→
- L3The price of a tokenAt depth four the grammar has sixteen leaves, and every local rule has been seen by sample ten. How many samples until every distinct full string has been seen?Measure, on a compositional grammar, the sample cost of learning surface strings against the cost of learning the local rules that generate them.→
- L2The copy sent inwardRotate the fly's head one hundred and eighty degrees, so self motion enters its eye with the wrong sign. What does the fly do?Build the efference copy: use a copy of your own motor command to separate what the world did from what you did, then break it the way von Holst broke it in 1950.→
The label that lies
Measure how an observer's reconstructed action labels absorb the substrate's work in proportion to stiffness, why the lie is invisible in calm air, and why it becomes the whole outcome the moment the body holds a load.
- L3The observer's physicsThe observer reconstructs actions with textbook physics that has no spring. Its label error, plotted against pole angle, has slope:Reconstruct actions from observed motion the way video pipelines do, and measure the systematic error the missing substrate leaves in the labels.→
- L4Benign in calm airThe labels carry a measured slope-sixteen systematic error. Clone from them and evaluate in calm air. Region drops by:Clone the same demonstrations twice, from true and from reconstructed labels, and measure the provably wrong labels costing nothing in calm conditions.→
- L4Fatal under loadAdd a steady 1.5 newton side-load, and retrain both clones on demonstrations recorded inside it. The reconstructed-label clone:Add one steady load, retrain both clones inside it, and measure the reconstructed-label clone collapsing from perfect to zero at a clocklike failure time.→