Revisited · 1935 Ripe now Robotics ✓ read
The Problem of the Interrelation of Coordination and Localization
Archives of Biological Sciences (Moscow; in Russian), 1935 ·not indexed by OpenAlex/Crossref ·7 min read
TL;DR. In 1935, Nikolai Bernstein argued that the brain cannot store movements as fixed motor commands, because the same efferent signal produces different movements depending on posture, load, and fatigue; instead, movement must be planned in task space and realized through flexible, low-dimensional synergies wrapped in fast sensory-correction loops. This is precisely the failure mode of today's flat end-to-end robot policies — token-space action outputs are muscle-space commands — and the latent action spaces and hierarchical controllers now appearing in robot foundation models are Bernstein's architecture being rediscovered piecemeal, without the theory. (Bibliographic details of the original Russian paper are not verified; the standard English source is the 1967 translation noted at the end.)
The idea as Bernstein had it
The reigning theory in 1935 was Pavlovian: movements are chains of conditioned reflexes, each stored somewhere specific in cortex — a punched tape of muscle commands, replayed on demand. Bernstein demolished this with two observations that any roboticist will recognize instantly.
First, the degrees-of-freedom problem. The human body has on the order of hundreds of muscles and dozens of kinematic degrees of freedom (the hand alone has roughly 20–27, depending on how you count). Any task constrains far fewer dimensions than the body offers. A controller that must specify every muscle independently faces a combinatorially hopeless credit-assignment problem. Bernstein's conclusion: the controller must not work at the muscle level. It must collapse the redundancy by binding degrees of freedom into task-shaped units — synergies — and control those.
Second, and more devastating for the punched-tape view: context-conditioned variability. The mapping from neural command to movement is not a function; it depends on state the brain does not fully know. Muscle force depends on current length and velocity. Joint torque from a given force depends on posture. Interaction torques couple every joint to every other. Spinal reflex gains modulate the whole thing. So an identical efferent signal yields different movements in different contexts — and, conversely, achieving the same movement requires different signals each time. Bernstein's famous later phrase for skilled practice, "repetition without repetition," follows directly: you can never replay a movement, only re-solve it.
His positive proposal: the stored representation of a movement — the "engram" — lives in an abstract, topological, task-space form (his evidence: your handwriting keeps its style whether you write small with your fingers, large on a blackboard with your shoulder, or clumsily with your foot). Execution is delegated to lower levels that translate task intent into muscle activity through synergies, with continuous sensory corrections closing the loop against the unpredictable periphery. Coordination is not localized; it is organized — a hierarchy of loops, not a library of tapes.
Why it could not work in 1935
Bernstein had the diagnosis but no instruments for the cure, on three fronts.
Measurement. His data came from kymocyclography: small light bulbs strapped to joints, photographed on moving film. He achieved impressively high frame rates for the era (on the order of 100+ exposures per second), but every trajectory had to be digitized by hand with a ruler, and differentiated by hand to get velocities and accelerations. One movement study was weeks of manual labor; a modern optical mocap rig gives sub-millimeter, hundreds-of-hertz, dozens-of-markers data in real time, and a decent EMG or high-density electrode array gives the muscle-space signal he could only infer. Roughly speaking, the cost of quantifying one reach has fallen by six-plus orders of magnitude in time and effort.
Theory and computation. Testing "flexible synergies with sensory corrections" means fitting state-feedback controllers and doing dimensionality reduction — regressions and eigendecompositions on thousands of samples. In 1935 there was no formal control theory to fit (Wiener's cybernetics is 1948; state-space methods and LQG are the 1960s–70s; stochastic optimal feedback control for motor behavior is Todorov & Jordan, 2002) and no machine to fit it with. A human computer manages perhaps one multiply-accumulate per second; even validating a single closed-loop model of a reach would have taken years of arithmetic. Bernstein could only argue verbally and geometrically — which is exactly what the paper does, brilliantly.
Politics. Bernstein's argument was a direct attack on Pavlovian orthodoxy. After WWII, and especially around the Pavlovian sessions of the early 1950s, his work was denounced and his laboratory activity curtailed; the ideas reached the West mainly through the 1967 Pergamon translation, decades late. (I'm confident of the broad outline; exact dates and institutional details vary across accounts.)
What changed, and what already vindicated him
Every pillar of the paper has since become a research program.
- Context-conditioned variability → optimal feedback control. Todorov & Jordan (2002) formalized exactly Bernstein's picture: don't replay trajectories, specify a task-space cost and let a feedback law resolve redundancy online. Their minimal intervention principle — correct deviations only in task-relevant dimensions, let task-irrelevant variance run free — explains "repetition without repetition" quantitatively, and the uncontrolled-manifold analyses of Scholz & Schöner confirmed that human movement variance is structured exactly this way.
- Synergies → measured. Santello and colleagues (1998) showed that when humans shape their hand to grasp dozens of objects, the first two principal components of hand posture capture roughly 80% of the variance: a ~20-DoF hand controlled through a ~2–3 dimensional synergy space. Bizzi, d'Avella and others found analogous low-dimensional muscle-activation modules in frog and human EMG.
- Robotics. Eigengrasps (Ciocarlie & Allen) used exactly those postural synergies to make grasp planning tractable. The Pisa/IIT SoftHand builds a synergy into the hardware — one motor driving a first-synergy tendon layout, with mechanical adaptivity playing the role of low-level correction. And today's robot foundation models are converging on the same shape from the learning side: latent action spaces learned from video (LAPA, Genie-style latent actions), action-chunking policies, VLAs at ~5–10 Hz sitting on top of whole-body or impedance controllers at hundreds of Hz. That stack is Figure 1's right-hand side — usually assembled ad hoc, without asking Bernstein's questions about what each layer should represent.
The failure mode he predicted is also visible: a flat policy emitting open-loop action tokens is a punched tape. It works until context shifts — a heavier object, a changed grasp, contact where none was expected — and then it fails in exactly the way a 1935 reflex-chain theory fails, because nothing in the loop is fast enough or task-aware enough to absorb the variability.
What a serious 2026 revival looks like
Take Bernstein literally and build a dexterous-hand (or humanoid whole-body) controller as an explicit three-level synergy hierarchy, then measure what the structure buys.
Architecture. Bottom: a fast loop at roughly 500 Hz–1 kHz over proprioception and tactile sensing, implementing minimal-intervention feedback — a learned or model-based OFC layer whose gains are conditioned on the current synergy. Middle: a learned low-dimensional synergy space, z \in \mathbb{R}^k with k \approx 3\text{–}10 for a hand, trained so that a decoder \pi(u \mid z, s) maps synergy plus proprioceptive state s to actuator commands u — crucially state-conditioned, so the same z yields different muscle-space output in different contexts, which is the whole point. Top: a VLA or planner at ~10 Hz that outputs trajectories in z and task-space goals, never touching actuators.
What to reuse from the paper: task-space engrams (the top level should represent goals topologically — object-relative frames — not joint trajectories); synergies as flexible mappings, not fixed PCA bases; corrections as first-class citizens, with the loop rate budget spent at the bottom.
What to replace: his verbal physiology with learned components; hand-designed synergies with latent spaces trained across tasks and, critically, across embodiments — his handwriting argument predicts the same z-trajectory should transfer between a five-finger hand and a three-finger gripper better than any joint-space policy does. That's a falsifiable, novel benchmark: motor equivalence as a transfer metric.
Evaluation: sample efficiency and robustness versus a flat end-to-end policy of matched capacity, under held-out perturbations (mass changes, surface friction, mid-grasp pushes) and cross-embodiment transfer. My expectation, from adjacent evidence, is a large win on perturbation robustness and transfer, a modest win on sample efficiency, and a possible loss on peak in-distribution performance — which would itself be an informative result about when synergy bottlenecks help versus hurt.
What is still open
Todorov himself raised the sharpest objection: synergies might be epiphenomenal — the low-dimensional structure could just be the shadow cast by optimal control of a low-dimensional task, not a real bottleneck in the controller. Whether synergies are an architectural commitment or an emergent statistic is still unresolved in neuroscience, and it maps onto a live robotics question: should latent action spaces be a hard bottleneck (Bernstein) or just a pretraining artifact that end-to-end fine-tuning is free to erode? Nobody has run that ablation cleanly at foundation-model scale. Also open: how new synergies are acquired during skill learning (Bernstein's later "levels of construction" story is qualitative), and how corrections are routed to the right level — the hierarchical credit-assignment problem, in modern language.
Where to read it
The 1935 paper (Russian original; bibliographic details not verified here) is available in English as a chapter of N. A. Bernstein, The Co-ordination and Regulation of Movements, Pergamon Press, 1967 — the essay usually titled "The problem of the interrelation of co-ordination and localization." Read alongside: Todorov & Jordan, "Optimal feedback control as a theory of motor coordination," Nature Neuroscience 2002 (the formalization); Santello, Flanders & Soechting, J. Neurosci. 1998 (synergies measured); Latash's writings on "Bernstein's problem" (the intellectual history); and any recent latent-action VLA paper — then notice that the 1935 essay already contains its block diagram, minus the GPUs.