Nociception as a Control Primitive: Afferent Channels and Nociceptive Memory for Agents Deployed in One Body

Wolfgang Maass

cs.AISep 28, 2026

Nociception as a Control Primitive: Afferent Channels and Nociceptive Memory for Agents Deployed in One Body

Authors: Wolfgang Maass

Organizations: Saarland University and German Research Center for Artificial Intelligence (DFKI)

Abstract

An agent deployed in a single body cannot learn how fast that body wears, because every trial that would reveal its wear resistance wears the body it would protect. We study this \emph{epoch-one} setting, in which the parameters of a fixed-weight policy are set before the body is drawn and never updated in life. The agent carries a load-gated nociceptive channel and a memory that retains what was felt. We prove that felt cost moves the allocation to the best-\emph{paid} work not yet felt rather than the gentlest, that an agent without retention never sees the felt-cost constraint bind, and that the channel pays only where the threat is individually unpredictable, cheap to avoid and expensive to ignore. We measure per body, setting the agent with channel and memory against the same individual without them, where neither carries a schedule learned across lives. On $2{,}000$ simulated floor-layer knees, with wear anchored to published loss rates, feeling, retaining and substituting extends the working life from age $55.2$ to $59.6$ and raises career output from $33.7$ to $36.1$ . $69.3\%$ of bodies gain and \textbf{none lose}. A body that feels but retains nothing past the day gains one of the $+4.4$ years, and retention carries the rest. A population-trained agent gains $+0.65$ years from the same channel at $-0.54$ output. The difference is what a species prior already supplies, and a single body has none. The two are related by an identity, the ablation mean reporting $(1-χ)$ of the per-body value with $χ$ the share a blind schedule already captures, so we report both. Where the regime map predicts value, a care robot sextuples its certified service life and a field-anchored fleet writes off $0.15$ of its machines instead of $0.55$ . Where it predicts none, a rover gains little over blind caution, so the map holds in both directions.

Figures & tables

Panel of Figure 1 — Figure 1: Architecture. The fast loop is one life inside a single hidden draw of the world. The dashed line is the observation boundary, crossed only by the afferent signal. The decision level allocates the day, which sets the load, the wear and the strength of the next signal, and its in-life state is the recorded felt cost per activity (probe) or a nociceptive memory (trained arm). The slow loop feeds the fast loop only before the life begins, with the draw $(\theta,e)$ and with the parameters of both the afferent layer and the decision level, and receives nothing back (Def. 1 ). The experimenter reads the whole life across the observation boundary and sets nothing inside it.

	floor layer, absolute				care robot		rover		fleet
what the body does	works to	output	$D_{\rm phys}$	$M_{\rm phys}$	life	output	life	output	life	output
works at full effort, reads nothing	$55.2$	$33.7$	$0.227$	$0.739$	$51.5$	$36.9$	$120.1$	$96.4$	$74.9$	$67.6$
spares from the first day †	$61.5$	$34.4$	$\mathbf{0.188}$	$\mathbf{0.780}$	$+226\%$	$+121\%$	$+151\%$	$+101\%$	$+15\%$	$-10\%$
fixed endowment $\alpha{=}2$ †§	$\mathbf{61.6}$	$35.3$	$0.191$	$0.775$	—	—	—	—	—	—
trained, feels nothing ‡	$59.2$	$33.4$	$0.200$	$0.768$	$+67\%$	$+30\%$	$-30\%$	$-46\%$	$-6\%$	$-26\%$
trained, full stack ‡	$59.9$	$32.9$	$0.191$	$0.778$	$+679\%$	$+432\%$	$+26\%$	$+2\%$	$+26\%$	$+20\%$

Table 1: Three ways to use one body, in four worlds, on the same individuals. All arms are paired, and no arm carries a schedule learned across lives (Def. 1 ). Floor layer: works to is the age the trade ends, output career output,

D_{\rm phys}

cartilage lost,

M_{\rm phys}

meniscal competence retained (

2{,}000

bodies). Machines: the first row gives the service life (certified months for the care robot, months otherwise) and the output of the arm that reads nothing (

500

paired lives per robot,

5{,}000

in the fleet), the other rows the percent change against it. The fleet runs to a horizon of

100

months, which the retaining arm reaches in every life, so its gain is a lower bound. † Not admissible as a reference, since it already knows which work is gentlest. § Trade only, since a machine reads every activity at once and an endowment is overwritten by the first reading. ‡ PPO. Trade:

10^{6}

environment steps,

24

paired seeds,

50{,}000

paired lives per seed, output as the trade paid it. Robots:

3\times 10^{6}

steps,

8

seeds, the rover at residual stream

0.10

. Fleet:

10^{6}

steps,

18

seeds, entropy

0

(App. C ). The blind trained arm already carries the population’s schedule (it reaches

59.2

without a channel), so these rows measure what a species prior gains from the channel, while the value to one body is the last row against the first.

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 2: The same stack in four worlds, both endpoints per world. Bars are the arms of Table 1 on the same paired bodies, and the dotted line is the body that reads nothing. In the trade and the care robot the retaining arm and the present-signal arm separate in opposite directions, slot (b) paying in one and costing in the other. In the rover blind caution reaches almost the whole gain, and in the fleet the one-shot memory alone takes the write-off rate from

57\%

0

(App. Table 7 ), while the present-signal arm stays within

1\%

of the reference, the one-step twitch of condition (b). Its trained arms are in Table 1 and App. Table 2 . Drawn from results/fig_arme_welten.json , which collects the files of Table 1 and App. C .

	wear takes the movement away		no wear
what the body carries	floor layer	care robot	latent-defect agent
	loses trade	decertified	written off
feels, minds, remembers	$\mathbf{1.8\%}$	$\mathbf{78.7\%}$	$\mathbf{17.8\%}$
feels nothing	$8.8\%^{***}$	$100\%^{**}$	$55.2\%^{***}$
feels, is not moved to act	$9.7\%^{***}$	$89.6\%$ n.s.	$18.8\%$ n.s.
feels and acts, but forgets	$1.2\%$ n.s.	$90.4\%$ n.s.	$53.5\%^{***}$

Appendix

Table 2: The same three parts, for trained policies. The share of lives that lose the paid work before the horizon, for one chain run through a floor layer’s cartilage, a care robot’s joints and a latent-defect fleet (

10^{6}

environment steps,

3\times 10^{6}

for the robot).

18

seeds per arm,

8

for the care robot,

{}^{***}p<0.001

and

{}^{**}p<0.01

against the intact body, n.s. marks a cell the measurement cannot separate from it.

variant	entr.	full	$-\,s$	$-\,h$	$-\,w$	ignorant	memory heur.
cascade $\times 14$	$0.05$	$0.38$ / $0.59$	$0.59$ / $0.50$	$0.55$ / $0.50$	$0.42$ / $0.59$	$0.68$ / $0.51$	$0.00$ / $0.83$
cascade $\times 14$	$0$	$0.15$ / $0.79$	$0.55$ / $0.49$	$0.53$ / $0.50$	$0.09$ / $0.79$
cascade $\times 39$ , $6$ seeds	$0.05$	$0.51$ / $0.50$	$0.59$ / $0.48$	$0.56$ / $0.48$	$0.53$ / $0.48$	$0.69$ / $0.49$	$0.00$ / $0.83$
$9\%$ defective, $\times 14$	$0.05$	$0.75$ / $0.913$	$0.76$ / $0.913$	$0.76$ / $0.908$	$0.76$ / $0.913$	$0.68$ / $0.80$	$0.00$ / $0.83$
$9\%$ defective, $\times 14$	$0$	$0.61$ / $0.930$	$0.76$ / $0.912$	$0.76$ / $0.910$	$0.65$ / $0.923$

Appendix

Table 3: Field-anchored fleet (Clusters 196499 and 196650,

18

seeds per cell unless stated,

2{,}000

paired lives per cell): death rate / welfare

J

, medians over seeds. For the

9\%

variant the death rate is conditional on carrying the defect and

J

is unconditional. Right: the same bodies under the ignorant static and the one-shot nociceptive-memory heuristic.

what was wrong	what it produced, and the check
The floor was switched off in both machine configurations	Sparing costs nothing, so the felt arms lose, $-11.7$ output against blind caution, and with the floor restored the same probe gains $+10.2$ months. Check that the floor can bind.
The rover’s floor was on the wrong quantity	It scored science yield while the mission is the traverse, so a portfolio parked at the instrument station cleared the floor at zero speed and the sparing arm completed $0$ of $3$ milestones in every life. Fixed with a progress floor at the rover’s own plan rate.
The rover’s mission ends long before its career does	Every policy completes all three milestones by month $55$ – $79$ of $960$ and earns the same reward to within $2\%$ , so $92\%$ of the horizon is driving with nothing left to achieve. No floor repairs this, and the run was withdrawn for the design in Table 1 .
The horizon censored the protected arms	$91\%$ of felt lives ran into the care robot’s horizon against $4\%$ of insensate ones, understating the effect roughly tenfold, $p{=}0.67$ becoming $p<10^{-30}$ once the horizon was lifted. Check that censoring falls on both arms alike.
Career length was not in the objective	Under average-reward training an episode ending on losing the work costs nothing, since episode length does not enter a per-step average, mean reward $0.143$ without an absorbing exit against $0.0012$ with it. Every ablation then beat the intact stack. One common functional across arms, and the training signal is never the endpoint.
Career length could not move in the capacity-exit acute knee	Every fixed probe reached exactly $65.00$ over $120$ bodies, since under maximal load damage reached only $D=0.157$ by $65$ and the only exit was the clock. Re-anchored, the exit fires in $41.4\%$ of lives and careers end at $60.7$ . Check, before any arm is compared, that the world can express protection at all.

Appendix

Table 6

Figure 3: Both headline numbers against the specialization axis. The band is the anchoring, not the sampling. Paired,

2{,}000

bodies per point. (a,b) Career length and output against the degree of specialization, the band spanning the anchored lost flexion

\Delta\theta\in[8^{\circ},27^{\circ}]

( Zhou et al., 2023 ) (a confidence interval would span

\pm 0.2

years). (c) Per-body distribution at the observed mix: filled, the whole apparatus (

69.3\%

gain, none lose); outline, the memory’s part alone (

65.1\%

gain, none lose), which carries

+3.4

of the

+4.4

years.

Figure 4: The middle link, over

2{,}000

paired bodies. Means over the bodies still working at that age, drawn while at least a fifth of them are. (a) The day’s share on the work that pays best and wears the joint. (b) How many activities the day contains at all, each at least

5\%

of it. (c) The share on the two gentle grips the trade offers (load below

20

). (d) The same day at

60

, heavy (dark) to gentle (pale), activities under

5\%

pooled at the right. Mechanism, not result: the result is printed below.

Figure 5: The chain, with the quantity that carries each link. Means over the

2{,}000

paired bodies of the anchored cell, the same bodies in every arm. Escape load is the hazard-weighted mean of the non-kneeling part of the day, and returns counts the years a body takes the heavy work back after leaving it. The two head numbers sit at the last link, and everything left of them is mechanism.

	the result		the mechanism
what the body carries	works to	output	on the knees	on gentle work	returns	$D$ at $50$
reads nothing	$55.2$	$33.7$	$0.70$	$0.00$	$7.3^{\dagger}$	$1.00$
present signal, no retention	$55.7$ – $56.2$	$33.5$ – $34.3$	$0.70$	$0.01$	$6.6$	$0.94$
reads the signal and remembers	$\mathbf{59.6}$	$\mathbf{36.1}$	$\mathbf{0.44}$	$\mathbf{0.36}$	$\mathbf{0.0}$	$\mathbf{0.84}$

Appendix

Table 5: The middle link as a quantity, on the same

2{,}000

paired bodies as Table 1 . Working life only. On the knees and on gentle work are shares of the day over the last five working years, gentle meaning the two grips below load

20

. Returns counts the years the heavy work is taken back after having been left.

D

50

is cartilage damage relative to the insensate body, over the

1{,}369

bodies still in the trade at

50

in every arm. Ranges are a tie-break span, not a confidence interval. † Not a decision: that share moves because the joint stops delivering. Source results/substitution_spur_b042.json .

cell ( $\Delta\theta$ , work available)	$V_{\rm ind}$ years	$95\%$ CI	$V_{\rm ind}$ output	$95\%$ CI
$8^{\circ}$ , $0.58$	$+1.47$	[ $+1.34$ , $+1.60$ ]	$\mathbf{-0.69}$	[ $-0.79$ , $-0.60$ ]
$8^{\circ}$ , $1.00$	$+1.70$	[ $+1.54$ , $+1.85$ ]	$+0.30$	[ $+0.18$ , $+0.41$ ]
$18^{\circ}$ , $0.58$	$+4.38$	[ $+4.23$ , $+4.53$ ]	$+2.43$	[ $+2.33$ , $+2.53$ ]
$18^{\circ}$ , $0.80$	$+5.32$	[ $+5.13$ , $+5.50$ ]	$+3.51$	[ $+3.38$ , $+3.63$ ]
$18^{\circ}$ , $1.00$	$+6.03$	[ $+5.81$ , $+6.25$ ]	$+4.26$	[ $+4.11$ , $+4.41$ ]
$27^{\circ}$ , $0.58$	$+4.81$	[ $+4.69$ , $+4.92$ ]	$+3.13$	[ $+3.05$ , $+3.21$ ]

Appendix

Table 6:

V_{\rm ind}

on both headline numbers, against a body that senses nothing and learns nothing (

2{,}000

paired bodies per cell, seven of fifteen cells, the rest interpolate monotonically). The second axis is how much of the trade’s work one worker may specialize into, and

0.58

is the observed population mix. All cells are the pinned world results/bodenleger_lp_b042 . Every interval excludes zero.

world (exit rule)	same body without / with	body endpoint	career (y, $95\%$ CI), and $J$	gain/lose
knee, wear with tears, $\Delta\theta{=}18^{\circ}$ (total-output floor, gate cell)	ignorant / full	early exit $0.69\to 0.04$ ; tears per exposed year unchanged	$+12.1$ ; $J$ $+11.6\%$ (all)	$0.66$ / $0.00$
knee, wear with tears, $\Delta\theta{=}8^{\circ}$	same	cap rarely binds	$+4.3$ ; $J$ $-2.8\%$	$0.26$ / $0.00$
fleet (absorbing death)	uniform static / one-shot nociceptive memory	written off in $57.1\%$ of lives against $0\%$ ; service $74.9\to 100.0^{\ddagger}$ months, output $67.6\to 90.2$	$J$ $+0.227$ $[0.220,0.234]$ ( $+36\%$ )	$0.57$ / $0.00$
care robot, anchored floor	full contact / reads the signal and spares	service $51.5\to 323.3$ months	output $36.9\to 150.8$ ( $+309\%$ )	$1.00$ / $0.00$

Appendix

Table 7:

V_{\rm ind}

across the worlds: the same body with and without a fixed protective response, no schedule learned across lives. Paired bodies: knee

20{,}000

per cell, fleet

5{,}000

, care

500

, and the rover’s probes are in App. C (

500

lives).

J

is the common functional of Eq. 2, all five random streams are seeded per life, and the care row is measured at the anchored floor with an uncensored horizon. ‡ Censored: see below.

Explore similar work

Jun 16, 2026cs.AI

Memory as a Wasting Asset: Pricing Flash Endurance for Embodied Agents, and the Limits of Doing So

A robot's flash endurance is a non-renewable stock: every persisted write spends one of a few thousand program/erase cycles and never refills, yet no fielded robot memory system prices which memories are worth an erase cycle. We treat embodied memory as depreciating capital and price that stock with a single endurance shadow price

η

, which makes cost-minimizing placement across a RAM / on-board NVM / cloud hierarchy a threshold in a wear-augmented per-byte index. The index is cost-optimal whatever the sign of the value-write association

χ

; only when

χ> 0

does the optimum turn non-monotone, sending a robot's most valuable memories off its flash. The pivot is thus empirical, and we measure

χ

on real robot logs at a pre-specified gate: its sign is a property of the deployment regime -- positive on recurrent long-horizon manipulation (

\hatχ \approx +1.0 \times 10^{-3}

, replicated at full power), null on a shorter-horizon suite, and negative on non-recurrent teleoperation. Two boundaries scope the result. The endurance budget is dormant on premium 3,000-P/E TLC at datasheet prices and binding on the commodity QLC/eMMC (

\sim

1,000 P/E) that cheaper edge robots run. And where it binds, a learned wear-aware controller only ties price-based routing on task value, because realized value is tier-invariant across RAM, NVM, and cloud: the rent governs device lifetime and cost, not task performance. Whether wear-aware placement improves task value remains open --

χ

is measured against a value proxy, and the non-monotone optimum, while proven, is not yet observed in data.

Josef Liyanjun Chen

KAIKAKU

Cost-Aware Embodied Agents Waste Wearable Devices

Ask PDF

Aug 4, 2026cs.AI

Interoceptive Attention as Dynamic Homeostatic Prioritization in a Foraging Agent

Biological systems must regulate competing needs under limited perceptual bandwidth, where sharpening one estimate costs the capacity to sharpen the others. Any fixed-budget system therefore has to decide where to allocate its perceptual precision. We study this in a foraging agent that must keep several bodily needs satisfied to survive, modelled with active inference. At each step it reads its own body-state beliefs, identifies the most-needed channel, and reallocates a fixed budget of interoceptive precision toward it, so that the same precision-shaped likelihood feeds both belief update and planning. In AffectWorld, a four-channel foraging gridworld, this selective allocation more than doubles learning-phase survival at matched budget against a uniform-precision agent (

0.414

0.199

across 11 layouts,

n{=}32

seeds each, paired cluster-bootstrap

p \leq 10^{-4}

). Two further results sharpen the mechanism. The benefit runs through planning as well as perception, since denying the shaped likelihood to the planner alone removes about half of it. It is also need-aligned, since aiming precision at the least-needed channel does worse than spreading it evenly. The attended channel additionally learns its own dynamics about twice as fast, and stays ahead even at matched observation count, a behavioural trace of the same precision routing, visible in learning speed, not survival.

St John Grimbly, Nicolas Kuske, Evert A. Boonstra +7

Dept. of Mathematics & Applied Mathematics, Univ. of Cape Town, South Africa · Neuroscience Institute, Univ. of Cape Town, South Africa · MIND Institute and CSAM, University of the Witwatersrand, South Africa +3

Predator-Prey Models Preference Learning Perception Pipeline

Ask PDF

May 26, 2025cs.AI

Shutdownable Agents through POST-Agency

Many fear that future artificial agents will resist shutdown. I present an idea - the POST-Agents Proposal - for ensuring that doesn't happen. I propose that we train agents to satisfy Preferences Only Between Same-Length Trajectories (POST). I then prove that POST - together with other conditions - implies Neutrality+: the agent maximizes expected utility, ignoring the probability distribution over trajectory-lengths. I argue that Neutrality+ keeps agents shutdownable and allows them to be useful.

Elliott Thornley

Strategic Agents Artificial Intelligence Agents Neutrality Proposal Trajectory Generation

Ask PDF