The Co-ordination and Regulation of Movements

TL;DR. In 1967, a Soviet physiologist published a book arguing that skilled movement is not the playback of stored trajectories but the online solving of a redundancy problem — mastering hundreds of degrees of freedom by organizing them into hierarchical synergies, and learning not by repeating a movement but by "repetition without repetition": repeatedly solving the same motor problem under varying conditions. Nothing in 1967 could compute this; robots of the era existed precisely because they avoided redundancy. Today's 40+ DOF humanoids, latent-action models, and hierarchical VLA policies are converging on Bernstein's design almost line by line — mostly without testing his sharpest, still-open prediction about how training data should be collected.

The idea as Bernstein had it

Nikolai Bernstein spent the 1920s–1940s in Moscow photographing workers with cyclogrammetry — small lamps attached to limbs, filmed at high frame rates, giving him millisecond-scale kinematics decades before motion capture. His most famous observation, from blacksmiths hammering a chisel: the hammer tip's trajectory was more reproducible across strikes than the trajectories of the individual joints producing it. The nervous system was evidently not controlling joints; it was controlling a task variable and letting the joints vary in whatever way solved the problem on that particular strike.

From this and much else he built two claims. (The 1967 Pergamon book, The Co-ordination and Regulation of Movements, is a posthumous English compilation of papers spanning decades; I haven't verified its exact contents chapter by chapter, but the core arguments below are well attested in the secondary literature.)

First, the degrees-of-freedom problem. The human body has on the order of 100+ kinematic degrees of freedom and several hundred muscles. Worse, the mapping from neural command to movement is context-dependent: the same muscle activation produces different torques depending on joint angle, and different movement depending on velocity, load, and the reactive forces flying around a multi-link body. There is no fixed lookup table from command to outcome. So the brain cannot store movements as commands. It must control a low-dimensional problem and close the loop against a high-dimensional, nonstationary plant.

Bernstein's answer was synergies — task-specific couplings that bind many degrees of freedom into a few controlled variables — organized into a hierarchy of levels of construction. In his scheme, roughly: a level of tone (postural/muscular background), a level of synergies (rhythmic, whole-body coordination patterns), a level of space (movements addressed to the external spatial field — reaching, locomotion toward targets), and a level of actions (sequenced, object-directed, meaningful behavior — tool use, praxis). Higher levels don't micromanage lower ones; they pass down compact task parameters and lower levels absorb the redundancy and the disturbances.

Second, "repetition without repetition." If commands don't map fixedly to outcomes, practice cannot be the stamping-in of a trajectory. Each repetition faces slightly different initial conditions, fatigue, load. What improves with practice is the ability to solve the motor problem — Bernstein's phrase — under variation. Skill acquisition, in his account, proceeds by first freezing degrees of freedom (novices stiffen, reducing the problem's dimension), then gradually freeing them and finally exploiting reactive dynamics rather than fighting them. Motor learning is the construction of levels of control, not the memorization of output.

Why it could not work then

Bernstein had no way to make any of this computational. The book's ideas predate cheap matrix algebra, let alone learning.

Compute. A top machine of the mid-1960s (an IBM 7090-class computer) delivered very roughly 10^5 operations per second and tens of kilobytes of core memory — filling a room. Just simulating the rigid-body dynamics of a 40-DOF humanoid at 1 kHz needs, roughly, millions of floating-point operations per second before you control anything, and Bernstein's scheme requires closing feedback loops at multiple timescales and learning the synergy structure from experience. That's many orders of magnitude out of reach.

Control theory. The tooling of the era was linear: root locus, LQR was brand new. There was no principled way to handle redundant, nonlinear, contact-rich dynamics. Redundancy was not a resource to exploit; it made the inverse problem ill-posed, which to a 1960s engineer meant "unsolvable."

Robotics. Unimate (deployed 1961) had roughly five hydraulic degrees of freedom and worked by recording and replaying joint positions from a magnetic drum — literally the stored-trajectory theory Bernstein was arguing against, implemented in hardware. Industrial robotics succeeded for decades by rigidly eliminating context-dependence: bolt the workpiece down, stiffen the arm, replay the tape. Bernstein's problem was avoided, not solved.

Learning. "Repetition without repetition" is, in modern terms, a claim about training-distribution design and about learning policies rather than trajectories. In 1967 there was no formal notion of a policy, no function approximation worth the name, no way to state the claim, let alone test it.

What changed

Every ingredient Bernstein lacked now exists, and most of his specific claims have been independently rediscovered.

Redundancy exploitation is now theory, not mystery. Todorov and Jordan's optimal feedback control work (2002) formalized "repetition without repetition" as the minimal intervention principle: an optimal controller corrects deviations only along task-relevant dimensions, letting variance accumulate in the task-irrelevant null space. Scholz and Schöner's uncontrolled manifold (UCM) analysis (1999) gave a way to measure this, and human data robustly shows the predicted structure: variance is large along the manifold of joint configurations that leave the task variable unchanged, small orthogonal to it. This is the blacksmith result, derived from first principles.

Synergies became extractable and buildable. Bizzi, d'Avella and colleagues found that frog and human muscle activity during natural behavior is well described by a handful of muscle synergies (low-rank factorizations of EMG). In robotics, dynamic movement primitives (Ijspeert, Schaal) and, more recently, learned latent action spaces — VQ-VAE action tokenizers, latent-action pretraining from video, skill embeddings from play data — all discover compact, reusable coordination variables from data. A humanoid whole-body policy trained in massively parallel simulation (thousands of environments on one GPU) has to compress its 40+ actuated DOF into something like synergies to learn at all.

The hierarchy got built. Current humanoid stacks — a slow vision-language planner on top of a mid-rate learned skill policy on top of a kilohertz impedance/whole-body controller — reproduce Bernstein's levels A through D with almost embarrassing fidelity (see figure). This wasn't designed from the book; it was forced by the same physics that forced Bernstein's conclusions. That convergence is itself evidence he was right.

What a serious 2026 revival would test

The parts of Bernstein that are vindicated are the hierarchy and the synergies. The part that is not yet properly tested in robot learning is repetition without repetition as a data-collection and training doctrine. Current practice is largely its opposite: behavior cloning from teleoperated demonstrations, where operators are often asked to demonstrate consistently, and losses penalize deviation from the demonstrated trajectory in joint or end-effector space.

A clean experiment:

1. Two datasets, matched budget. Task: a contact-rich manipulation or loco-manipulation problem (e.g., insert a plug, carry an awkward box upstairs). Dataset R ("repetition"): N demonstrations of one nominal instance, demonstrated as consistently as possible. Dataset P ("problem-solving"): N demonstrations spanning varied object poses, masses, frictions, starting postures — including recoveries from perturbations mid-demonstration.

2. Two losses. Trajectory loss: standard behavior cloning in action space. UCM-inspired loss: penalize error in task variables tightly, but penalize null-space deviation weakly or not at all — formally, if J is the task Jacobian, weight action errors by projection onto \mathrm{range}(J^\top) and discount the null-space component. The policy is graded on solving the problem, not on reproducing the tape.

3. Test out of distribution: novel poses, added payload, mid-episode pushes. Bernstein predicts P + UCM-loss beats R + trajectory-loss decisively on OOD robustness, and predicts the trained policy's own trial-to-trial variance will show UCM structure (large null-space variance) — a measurable signature, not just a benchmark number.

A second Bernstein-specific prediction worth testing: freezing-then-freeing as a curriculum. Train the latent action space with an initially low-dimensional bottleneck (few synergies, high stiffness at the impedance layer), then progressively widen the bottleneck and lower stiffness. His account of human skill acquisition says this ordering matters; curriculum experiments could confirm or kill it.

What to reuse from the book: the level decomposition, the insistence that each level closes its own sensory loop (he called movement a "reflex ring," not an arc — feedback all the way down), and the variability doctrine. What to replace: everything neurophysiological. His anatomical assignments of levels to specific brain structures (rubro-spinal, pallidal, pyramidal) reflect 1940s neurology and shouldn't be taken literally.

Status, caveats, open questions

Honest caveats. First, the muscle-synergy literature has a live critique (Kutch, Valero-Cuevas, and others): low-rank EMG structure can emerge from task and biomechanical constraints without any neural synergy module, so "synergies exist as descriptions" is solid while "synergies are the controller's building blocks" is contested. Bernstein's engineering claim survives either way — for a robot, it doesn't matter whether biology implements synergies explicitly, only whether low-dimensional action spaces help. Second, end-to-end training keeps beating hand-imposed structure in ML; explicit Bernstein levels may lose to a flat transformer given enough data. The interesting question is data-efficiency and OOD behavior at realistic robot-data scales, where structure should matter most. Third, some evidence already points Bernstein's way — play-data and RL-finetuned policies generalize better than narrow demonstration cloning — but I'm not aware of a controlled study that isolates demonstration variability under matched budgets. That experiment is cheap and overdue.

Where to read it

The book is The Co-ordination and Regulation of Movements, Pergamon Press, 1967 — long out of print; find it via libraries or the substantial excerpts reproduced in Latash's edited volumes (Dexterity and Its Development, with Bernstein's more accessible late manuscript, is the friendlier entry point). Read alongside: Todorov & Jordan, "Optimal feedback control as a theory of motor coordination" (Nature Neuroscience, 2002) — the mathematical vindication; Scholz & Schöner's UCM paper (1999); d'Avella, Saltiel & Bizzi on muscle synergies (2003); and, on the robotics side, Ijspeert et al. on dynamic movement primitives and any recent latent-action pretraining paper — then look again at the levels diagram above and note who drew it first.