5D Library › Embodied Cognition, Bernstein's Degrees-of-Freedom Problem and the Control Theory of Motor Synergies · 中文
Chapter 4 Internal Forward Models and Efference Copy: Getting Around Sensory Delay
How long are the delays?
The models of the previous chapter all assumed that the controller always knows where the limb is. The real nervous system does not: transduction in the receptors, nerve conduction and central processing all take time, and by the time a sensory signal arrives the limb has already moved some distance.
Evidence
Delays in sensorimotor loops. The 2011 review of Franklin and Wolpert in Neuron states that delays in sensorimotor loops are of the order of 100 ms but depend on the sensory modality, vision being slower than proprioception: the monosynaptic stretch reflex takes about 10–40 ms, the cortical component of the long-latency stretch reflex adds about 20–30 ms, involuntary motor responses to visual stimuli take about 110–150 ms, and the vestibulo-ocular reflex about 10 ms (Franklin and Wolpert 2011).
So the framework’s “50–200 ms” can only be taken as a rough range: the fastest spinal reflexes are below 50 ms, corrections through vision often take more than 100 ms, and conscious reaction times are longer still. Book 29 discussed the pitch-shift reflex in auditory feedback, at about 100–150 ms, and how delay makes high-gain feedback control oscillate; this chapter does not repeat that control program and deals only with state estimation: when signals arrive late, how does the nervous system know where the limb is now?
Internal models
One direct solution is prediction. If the nervous system keeps a copy of the motor commands it sends, together with a model of how the body and the environment respond to commands, it can compute the consequences of an action before the sensory signals arrive. The former is called the efference copy, the latter the forward model; by contrast, a model that works back from the desired result to the command needed is called an inverse model (Kawato 1999; Franklin and Wolpert 2011).
Evidence
Where the two terms come from. In 1950 von Holst and Mittelstaedt proposed the “reafference principle”: when the centre issues a command it leaves an “efference copy”, which is compared with the sensory signals caused by the movement (the reafference), so that self-caused and externally caused sensations can be told apart (Holst and Mittelstaedt 1950). In the same year Sperry, studying the circling behaviour of fish whose eyes had been surgically inverted, proposed the concept of “corollary discharge” (Sperry 1950). The two are related but not identical: the first stresses a copy of the command, the second refers more generally to central signals accompanying motor commands. The framework writes “伴随放电 (Efference Copy)”; this book distinguishes “efference copy” (传出副本) from “corollary discharge” (伴随放电).
In 1993 Miall and colleagues proposed that the cerebellum may work like a Smith predictor from control engineering: a fast forward model predicts the sensory consequences of the action, a second model simulates the delays in the loop and holds the prediction back so that it arrives together with the real feedback, and the difference between the two is used to correct the action and to train the first model (Miall et al. 1993). The 1998 review of Wolpert, Miall and Kawato discussed further evidence for forward and inverse models in the cerebellum (Wolpert et al. 1998). The framework attributes “cerebellar internal forward models” to Wolpert and colleagues 1995; this needs correcting, since that paper studied an internal model in sensorimotor integration and did not concern the cerebellum.
Evidence
Wolpert, Ghahramani and Jordan 1995. Participants moved their arm in the dark, sometimes against externally imposed assistive or resistive forces, and estimated the position of their hand at the end of the movement. The authors analysed how the estimation error evolved over time within a state-estimation framework, and concluded that the results directly support the view that the central nervous system internally simulates the dynamic behaviour of the motor system — that is, that an internal model exists (Wolpert et al. 1995). Their model combines the forward model’s prediction with sensory feedback, each weighted by its uncertainty, which is the structure of the Kalman filter (Kalman 1960). The optimal multisensory integration discussed in Book 34 is the same principle applied across senses.
Kalman estimation: predict, then correct
The program below compares three ways of estimating the position of the hand. A one-dimensional “hand” (mass 1 kg, viscosity 2 N s/m) moves 0.15 m back and forth at 0.5 Hz for 2 s under a known command (the efference copy), with motor noise of SD 0.5 N at every step; proprioception reports position with SD 5 mm but steps late (10 ms per step). The first method uses only the latest (late) reading; the second only integrates the command through the forward model (dead reckoning); the third is a Kalman filter on a state augmented with the current position, velocity and the positions of the past steps, which at every step predicts with the forward model and corrects with the late reading. In a second condition the hand is acted on by an external force of +0.3 N that the forward model does not know about.
Calculation
# State estimation with an internal forward model and delayed feedback (illustrative).
# A 1-D hand (mass 1 kg, viscosity 2 N s/m) moves back and forth for 2 s under a known
# command (efference copy). Proprioception reports position d steps late (dt = 10 ms)
# with SD 5 mm. Motor noise: SD 0.5 N per step. In the "bias" condition an external
# force of +0.3 N acts that the forward model does not know about. Three estimators:
# the latest delayed reading alone; the forward model alone (dead reckoning); and a
# Kalman filter on the delay-augmented state [p, v, p(t-1), ..., p(t-d)].
import math, random
DT, M, VISC, SY, SU, STEPS = 0.01, 1.0, 2.0, 0.005, 0.5, 200
def command(t): # 0.15 m back-and-forth at 0.5 Hz, plus viscosity
w = 2 * math.pi * 0.5
return M * (-0.15 * w * w * math.sin(w * t)) + VISC * 0.15 * w * math.cos(w * t)
def model(d):
n = 2 + d; A = [[0.0] * n for _ in range(n)]
A[0][0] = 1; A[0][1] = DT; A[1][1] = 1 - DT * VISC / M
A[2][0] = 1
for i in range(3, n):
A[i][i - 1] = 1
return A, [0.0, DT / M] + [0.0] * d
def kalman_gains(d):
A, b = model(d); n = len(A); P = [[0.0] * n for _ in range(n)]; gains = []
q = (SU * DT / M) ** 2 * 4 # process noise, inflated for unknown forces
for _ in range(STEPS):
P = [[sum(A[i][k] * sum(P[k][l] * A[j][l] for l in range(n)) for k in range(n))
for j in range(n)] for i in range(n)]
P[1][1] += q
s = P[n - 1][n - 1] + SY ** 2; K = [P[i][n - 1] / s for i in range(n)]
P = [[P[i][j] - K[i] * P[n - 1][j] for j in range(n)] for i in range(n)]
gains.append(K)
return gains
def run(d, gains, bias, rng):
A, b = model(d); n = len(A); x = [0.0] * n; z = [0.0] * n; fm = [0.0, 0.0]
err = {"delayed reading": [], "forward model": [], "Kalman": []}
for t in range(STEPS):
u = command(t * DT)
noise = rng.gauss(0, SU)
x = [sum(a * v for a, v in zip(r, x)) + bi * (u + noise + bias)
for r, bi in zip(A, b)]
y = x[n - 1] + rng.gauss(0, SY)
z = [sum(a * v for a, v in zip(r, z)) + bi * u for r, bi in zip(A, b)]
inn = y - z[n - 1]; z = [zi + k * inn for zi, k in zip(z, gains[t])]
fm = [fm[0] + DT * fm[1], fm[1] + DT * (u - VISC * fm[1]) / M]
err["delayed reading"].append(y - x[0]); err["forward model"].append(fm[0] - x[0])
err["Kalman"].append(z[0] - x[0])
return err
if __name__ == "__main__":
rng = random.Random(1995)
print("RMS position error over the 2 s movement (mm), 300 runs; no external force")
print(" delay delayed reading forward model Kalman")
for d in (5, 10, 20):
g = kalman_gains(d); acc = {}
for _ in range(300):
for k, v in run(d, g, 0.0, rng).items():
acc[k] = acc.get(k, 0.0) + sum(e * e for e in v) / STEPS / 300
print(f" {10 * d:4d} ms {1000 * math.sqrt(acc['delayed reading']):6.2f}"
f" {1000 * math.sqrt(acc['forward model']):6.2f}"
f" {1000 * math.sqrt(acc['Kalman']):6.2f}")
print("unknown external force +0.3 N, delay 100 ms: mean error (bias) and SD, mm")
g = kalman_gains(10); runs = [run(10, g, 0.3, rng) for _ in range(300)]
print(" time forward model: bias SD Kalman: bias SD")
for t in (25, 50, 100, 150, 199):
out = []
for k in ("forward model", "Kalman"):
v = [r[k][t] for r in runs]; m = sum(v) / len(v)
s = math.sqrt(sum((e - m) ** 2 for e in v) / (len(v) - 1))
out.append(f"{1000 * m:+8.2f} {1000 * s:6.2f}")
print(f" {(t + 1) * DT:4.2f} s {out[0]} {out[1]}")
# RMS position error over the 2 s movement (mm), 300 runs; no external force
# delay delayed reading forward model Kalman
# 50 ms 17.02 16.46 2.40
# 100 ms 32.37 17.54 3.18
# 200 ms 61.84 17.05 4.57
# unknown external force +0.3 N, delay 100 ms: mean error (bias) and SD, mm
# time forward model: bias SD Kalman: bias SD
# 0.26 s -8.37 3.10 -5.20 3.18
# 0.51 s -28.33 7.38 -4.83 3.36
# 1.01 s -86.44 15.94 -5.02 3.32
# 1.51 s -155.26 22.98 -5.18 3.08
# 2.00 s -227.25 28.47 -5.61 3.24Without the external force, the error of the late reading alone grows in proportion to the delay: 17.02, 32.37 and 61.84 mm at 50, 100 and 200 ms, essentially the distance the hand moves during the delay. The forward model alone errs by about 17 mm regardless of delay, from accumulated motor noise. The Kalman estimate combines the two and errs by 2.40, 3.18 and 4.57 mm, about one seventh, one fifth and one quarter of the better single method (ratios 6.9, 5.5 and 3.7); the error grows with delay, but much more slowly.
With the unknown external force (delay 100 ms), the bias of the forward model alone keeps accumulating: −8.37 mm at 0.26 s, −86.44 mm at 1 s, −227.25 mm at 2 s. The bias of the Kalman estimate settles at about −5 mm (−4.83 to −5.61 mm) with an SD of about 3.1–3.4 mm: sensory feedback keeps pulling the estimate back, but as long as the model does not know about the force a steady bias remains (Figure 4, left). Removing it requires the model itself to learn the force, that is, to adapt; this is why internal models need continual recalibration.
Why you cannot tickle yourself
Another use of the forward model is to tell self from world. If the sensory consequences of one’s own action can be predicted, they can be subtracted from the sensory signal, leaving only the part caused by the outside world.
Evidence
The tickle studies. In 1971 Weiskrantz, Elliott and Darlington reported preliminary observations on “tickling oneself”: stimuli applied by oneself feel weaker than the same stimuli applied by someone else (Weiskrantz et al. 1971). In 1998 Blakemore, Wolpert and Frith used fMRI in six participants to compare self-produced and externally produced touch (a foam-tipped rod moved across the left palm by the participant’s right hand or by the experimenter): self-produced touch reduced activity in bilateral secondary somatosensory cortex and the anterior cingulate, and the right anterior cerebellum was deactivated by movements that produced touch and activated by external touch, leading the authors to propose that the cerebellum takes part in predicting the specific sensory consequences of movements (Blakemore et al. 1998). In the 1999 psychophysical experiment, movements of the left hand produced touch on the right palm through a robot, with delays of 0, 100, 200 or 300 ms inserted between them, or with the trajectory rotated by 30°, 60° or 90°; as the delay increased from 0 to 200 ms and the rotation from 0° to 90°, “tickliness” ratings rose progressively (Blakemore et al. 1999).
Evidence
Force escalation. Shergill, Bays, Frith and Wolpert in 2003 had six pairs of participants take turns pressing on each other’s finger, each instructed to apply the same force they had just felt, starting from 0.25 N. The force rose by a mean of 38 % per turn, by 3.2 N over eight turns, and escalation occurred in every pair. Twelve further participants were tested alone: when matching an external force by pressing with their own finger they systematically pressed too hard (compared with matching via a joystick, the intercept rose by 0.53 N on average and the slope by 49 %). The authors concluded that self-generated forces are predictively attenuated, so that each person feels the other pressing harder than they themselves did (Shergill et al. 2003).
Calculation
The first part of the program uses an illustrative model to show how delay and rotation weaken “cancellation”: the robot strokes along a two-dimensional path (randomly combined sinusoids from 0.5 to 3 Hz), the forward model predicts the touch with no delay and no rotation, cancellation is proportional to the match (normalised inner product) between prediction and actual touch, perceived intensity , averaged over 200 random paths. The second part assumes that each person feels their own force multiplied by an attenuation factor and next applies “the force I just felt, divided by ”; a mean escalation of 38 % per turn corresponds to .
# Predictive attenuation of self-produced touch (illustrative model). The left hand
# moves a robot that strokes the right palm along a 2-D path s(t): band-limited motion,
# sum of sinusoids 0.5-3 Hz with random phases (seeded). The forward model predicts the
# touch from the efference copy with no delay and no rotation; the robot may add a
# delay and rotate the path. Cancellation is proportional to the match rho (normalised
# inner product of prediction and actual touch): perceived = 1 - g * max(0, rho), with
# g = 0.8, averaged over 20 s and 200 random paths. Part 2: force escalation
# when each person feels their own push attenuated by a constant factor.
import math, random
def path(rng, n=2000, dt=0.01):
comps = [(f, rng.uniform(0, 2 * math.pi), rng.uniform(0, 2 * math.pi), 1 / f)
for f in (0.5, 0.8, 1.2, 1.7, 2.3, 3.0)]
return [(sum(a * math.sin(2 * math.pi * f * k * dt + px) for f, px, _, a in comps),
sum(a * math.sin(2 * math.pi * f * k * dt + py) for f, _, py, a in comps))
for k in range(n)]
def perceived(p, lag, rot, g=0.8):
c, s = math.cos(rot), math.sin(rot); dot = na = npr = 0.0
for k in range(lag, len(p)):
ax, ay = p[k - lag]; ax, ay = c * ax - s * ay, s * ax + c * ay # actual touch
px, py = p[k] # prediction
dot += ax * px + ay * py; na += ax * ax + ay * ay; npr += px * px + py * py
return 1 - g * max(0.0, dot / math.sqrt(na * npr))
if __name__ == "__main__":
rng = random.Random(1998); paths = [path(rng) for _ in range(200)]
print("perceived intensity relative to external touch (1.00 = no attenuation)")
print(" rotation: 0 deg 30 deg 60 deg 90 deg")
for ms in (0, 100, 200, 300):
vals = [sum(perceived(p, ms // 10, math.radians(r)) for p in paths)
/ len(paths) for r in (0, 30, 60, 90)]
print(f" delay {ms:3d} ms " + " ".join(f"{v:.2f}" for v in vals))
print("force escalation: next push = own previous felt force / attenuation factor")
for a in (1.0, 0.85, 1 / 1.38):
f = [0.25]
for _ in range(8):
f.append(f[-1] / a)
print(f" factor {a:.3f}: force after turns 1, 4, 8 = {f[1]:.2f}, {f[4]:.2f},"
f" {f[8]:.2f} N; rise over 8 turns {f[8] - f[0]:.2f} N")
rng = random.Random(2003); ends = []
for _ in range(10000): # 38 % mean per turn with 15 % turn noise
f = 0.25
for _ in range(8):
f *= 1.38 * math.exp(rng.gauss(0, 0.15) - 0.15 ** 2 / 2)
ends.append(f - 0.25)
ends.sort()
print(f" factor 0.725 with 15 % noise per turn: median rise {ends[5000]:.2f} N,"
f" 10-90 % {ends[1000]:.2f} to {ends[9000]:.2f} N")
# perceived intensity relative to external touch (1.00 = no attenuation)
# rotation: 0 deg 30 deg 60 deg 90 deg
# delay 0 ms 0.20 0.31 0.60 1.00
# delay 100 ms 0.32 0.41 0.66 0.94
# delay 200 ms 0.58 0.63 0.77 0.90
# delay 300 ms 0.81 0.82 0.83 0.86
# force escalation: next push = own previous felt force / attenuation factor
# factor 1.000: force after turns 1, 4, 8 = 0.25, 0.25, 0.25 N; rise over 8 turns 0.00 N
# factor 0.850: force after turns 1, 4, 8 = 0.29, 0.48, 0.92 N; rise over 8 turns 0.67 N
# factor 0.725: force after turns 1, 4, 8 = 0.34, 0.91, 3.29 N; rise over 8 turns 3.04 N
# factor 0.725 with 15 % noise per turn: median rise 2.76 N, 10-90 % 1.52 to 4.92 NIn this model, with no delay and no rotation the perceived intensity is only 0.20 of external touch; with delays of 100, 200 and 300 ms it rises to 0.32, 0.58 and 0.81; with a 90° rotation it is 1.00 (no cancellation at all). This agrees in direction with what Blakemore and colleagues reported in 1999, but the values depend on the assumed path frequencies and gain and are not a fit to the experiment (Figure 4, right).
In the escalation part, with factor 0.725 the force after eight turns is 3.29 N, a rise of 3.04 N, close to the reported 3.2 N; with factor 0.85 it rises by only 0.67 N. Adding 15 % multiplicative random variation per turn gives a median rise of 2.76 N, with 10th to 90th percentiles of 1.52 to 4.92 N. A constant self-attenuation is thus enough to produce the escalation seen in the experiment; this does not prove that the factor is constant, only that the reported numbers are mutually consistent in this simple model.
Hypothesis · not established · An internal focus amplifies the cancelled signal
The author’s conjecture: if the nervous system predicts and attenuates self-produced sensation with a forward model, then asking learners to attend to the sensations of their own joints and muscles (the “internal focus” of Chapter 6) may amount to raising the weight of signals that ought to be attenuated, making small prediction errors seem important and triggering unnecessary corrections — including corrections in task-irrelevant directions. What would refute it: within the same participants, internal-focus and external-focus instructions produce no difference in the over-estimation of self-produced force in a force-matching task (the measure of Shergill and colleagues), or no difference in the frequency of corrective submovements and the amount of correction in the UCM direction.
Summary
Delays in sensorimotor loops differ by modality: about 10–40 ms for the monosynaptic reflex and about 110–150 ms for visual responses, so “50–200 ms” is only a rough range. A forward model predicts the consequences of an action from the efference copy, and Kalman estimation merges the prediction with late feedback according to their uncertainties: in the illustrative model with a 100 ms delay the error falls from 32.37 mm (late reading alone) to 3.18 mm; an unknown external force leaves a steady bias of about 5 mm that only adaptation of the model itself can remove. The 1995 paper of Wolpert and colleagues concerns an internal model in sensorimotor integration; the cerebellar internal model goes back to Miall and colleagues 1993 and Wolpert and colleagues 1998. Predictive attenuation explains why you cannot tickle yourself (the fMRI of Blakemore and colleagues 1998 and the psychophysics of 1999) and the escalation of force (38 % per turn, corresponding to an attenuation factor of about 0.725). The next chapter turns from joints and senses to muscles: how muscle synergies were “discovered”, and whether they really come from the nervous system.
References
Fan, Y. (2026). Embodied Cognition, Bernstein's Degrees-of-Freedom Problem and the Control Theory of Motor Synergies: From the Uncontrolled Manifold Hypothesis and Internal Forward Models to Fitts's Law, Read Beside Classical Chinese Thought on Skill and the Way [具身认知、伯恩斯坦自由度问题与运动协同的控制理论] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23828057 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).