Sensing · energy · measurement
What a Machine Spends Looking
Task energy on a body is reported as compute against everything else. Sensing is measurable on the same meter, is not separated, and how large it has to be is set by the body before any perception exists.
1. What is in scope, and what does this continue?
One term of a sum that three companion reports have already established. TR-2026-40 §10 names the denominators a figure can be stated in, and shows that on a body the second of them decomposes: compute is one term, and improving it by a factor leaves the rest untouched, so the limit of that improvement is the reciprocal of the non-compute share. Measured, that share moves a long way: five percent of a reaching task at one watt of compute, fifty-nine percent at thirty watts, and on a mobile robot under autonomous navigation a graphics processor drawing 37.3 percent against 16.6 percent for the motors.
TR-2026-36 establishes where a meter has to sit before any of that is attributable, and grades ISO/TS 25213, the test method for the energy consumption of 6-axis articulated industrial robots, at stage 60.00 and effective 26 June 2026 catalogue record and scope verified; clause text not read. TR-2026-35 reads the transfer literature across twenty-three regions in eighteen languages, and TR-2026-43 measures the same body property this report sweeps, against demonstration volume rather than against a sensor set. The viability denominator this Institute reports against, joules per viability-second held under disturbance, is already demonstrated across six bodies including a physical SO-101 arm at 140 randomised seeds.
This report adds one term and one body property. The established split is compute against non-compute. Sensing sits inside the second half, undifferentiated, and it is separable on the same instrument. And the size of the sensing bill turns out to be set by a mechanical property that is fixed long before a perception stack exists.
2. What does a machine spend looking?
A term nobody separates, on an instrument that can already separate it. A sensor read costs energy in three places: the transducer's own draw, the conversion and transport of the sample, and the computation that consumes it. The first two are properties of the channel and are constant per read, which makes them straightforwardly meterable on a dedicated rail alongside compute. What makes the term worth separating is not its size but its controllability: unlike actuation, which a disturbed body must spend, and unlike a policy's compute, which is fixed by the architecture, the sensing bill is set by a schedule the designer chooses.
On the bench used here, four channels carry per-read costs of 0.0090, 0.0060, 0.0030 and 0.0025 joules for cart position, cart velocity, pole angle and pole rate. Priced against the region each buys, the attitude pair costs 2.20 J and buys its region at about 19.3 J per point; the position pair costs 6.00 J and buys it at about 164. The protective channels are cheaper in absolute terms and by about 8.5x per unit of viability bought, which is what makes an always-on protective layer affordable while continuous perception is not.
3. What sets the size of that bill?
The body, and it is decided in mechanical design. Sweeping the pivot stiffness of a cart-pole with every sensor switched off and the controller byte-identical in every row, the share of the episode held inside the barrier moves from 0.000 at zero stiffness to 0.360 at 0.50, 0.794 at 1.00 and 0.834 at 2.00 N·m/rad. Nothing about the policy changed. What changed is whether the upright is a stable equilibrium of the mechanism.
The threshold is a property of the body's own equations rather than of the textbook form. The gravitational stiffness of a pendulum about a fixed pivot is m g L = 0.4905 N·m/rad for this body, but the pole sits on a free cart whose recoil takes angular acceleration with it. Linearising the implemented step at the upright gives m g L · (4/3) / (4/3 − mp/(mc+mp)) = 0.5264, 7.3 percent higher, confirmed by bisection on the discrete step at 0.5263. A design that labels stiffness against the fixed-pivot figure calls an unstable body stable across part of its range.
On a second body the same law holds, and the plateau sits higher. SENSE/2 runs the same sweep on a two-link standing body, ankle and hip, where the threshold is the largest eigenvalue of the coupled gravitational stiffness matrix at 51.25 N·m/rad, 1.13x the largest single-joint figure. The no-sensor region rises to 0.682 at that threshold, 0.984 by 1.5 times it and 1.000 by twice it, flat out to 2.7 times. An earlier version of this section reported that same curve peaking at the threshold and collapsing to zero past it, and concluded that the cart-pole's plateau was peculiar to its morphology. That collapse was the integrator rather than the body, and both halves are withdrawn in §7.
The complementary half is what sensing then buys. At a stiffness of 1.0 the blind loop holds region 0.794 while its task error sits at 0.936, indistinguishable from not attempting the task; the full four-channel suite takes task error to 0.487 and region to 1.000. Sensing bought the task, not the survival.
4. Why does more data not fix it either?
Because a dataset is a recording of sensor streams, so the two bills have the same floor. TR-2026-43 ran the same dial against a different resource on the same plant. Where §3 above sweeps pivot stiffness against a sensor set, that report sweeps it against a demonstration count, cloning a policy from camera observations the way a video pipeline sees the world. At maximum data, stiffness buys +0.742 of certified operating region. At zero stiffness, 256 times more data buys −0.027, flat within two standard errors. Two demonstrations on a compliant body certify the full region; five hundred and twelve on a stiff one reach 0.258.
Set beside §3 the two results say one thing. Compliance moves the region a blind controller holds from 0.000 to 0.834, and it moves the region a data-trained policy holds by +0.742, and in both cases the resource being scaled buys almost nothing without it. That is not two findings that agree. It is one finding reached from two directions, and the reason they cannot come apart is that a dataset is a recording of sensor streams. Torque exchanged directly between a compliant body and its load never crosses a transducer, so it is absent from the sensor stream in real time and absent from every recording of that stream, permanently.
The consequence is a shared floor, which neither report states alone. The sensing bill and the data bill are not two budgets to be traded against each other. They are two ways of paying for the same missing information, and both are bounded below by the same mechanical property. A programme that responds to a thin sensor budget by collecting more demonstrations is buying the second copy of what it already could not see. The dial that moves both sits in the mechanism, and it is set before either budget is written.
5. Does the second morphology agree?
On the threshold and the shape yes, on what sensing buys no. A result from one body is a data point, so §3's sweep was repeated on a two-link standing body: two coupled rotations rather than a rotation and a translation, inertial coupling between the links, and no free base to absorb momentum. The controller, the estimator, the disturbance, the seed schedule and the four sensor costs are held identical.
| Claim | Cart-pole | Two-link stand |
|---|---|---|
| The threshold is a coupled property, not a textbook figure | 0.5264 against the fixed-pivot 0.4905, 1.07x | 51.25 against the largest single joint 45.47, 1.13x |
| Compliance moves the no-sensor region | 0.000 to 0.834, monotone, flat from 2.0 out to 32 | 0.000 to 1.000, monotone, flat from 2x the threshold |
| What sensing buys | the task: error 0.936 to 0.487, region already held | the survival: region +0.323, task error only +0.014 |
Two of the three carry across and one inverts. On the cart-pole the blind body survives and cannot do the job; on the two-link stand at its threshold stiffness the blind body is marginal at 0.677, and the sensors are what hold it inside the barrier. Which of the two vectors a sense serves is a property of the body, not of the sense.
The two plateaus differ, and the difference is a second barrier rather than a disagreement. The cart-pole levels off at 0.834 and the stand at 1.000. Decomposing which barrier the blind cart-pole actually hits accounts for the gap. At a stiffness of 0.5 every failure is an attitude failure. At 1.0 it is 70 percent cart speed. From 2.0 out to 32, not one failure is an attitude failure: all of them are the cart running past its velocity bound. The compliant pivot has removed the falling problem completely and left a base-excursion problem in its place. Giving the base authority closes the gap, with the policy still unchanged: at 4 kg the plateau is 0.944 and at 16 kg it is 1.000, matching the two-link stand, whose base is fixed to the ground and has no velocity bound to violate. Stiffening a joint converts an attitude failure into a base failure, and whether that is progress depends on whether the base has somewhere to put the momentum.
The sharper number is in the sensor sets. The ankle pair alone, at 2.20 J, reaches region 1.000 and task error 0.069: indistinguishable from the full four-channel suite at 8.20 J. The hip channels buy nothing measurable on this body. That is the same shape as §2's protective-against-generative pricing, at 3.7x the saving, and it is only visible because the channels were priced separately.
6. Which hypotheses does this test?
Four, stated before the bench existed, and one survived.
| # | Hypothesis | Resolution |
|---|---|---|
| H1 | The zero-sensor certified region is nonzero | Survives, and sharpens. It is a property of the body: 0.000 at zero pivot stiffness, 0.834 at 2.00, policy unchanged. |
| H2 | Sense value is strongly sublinear; the first sense buys most of the region | Falls. Not sublinear and not monotone: the second channel buys more region than the first (+0.053 then +0.061), so value here is synergistic rather than additive and a per-sensor ranking can mis-order the set to buy. Whether attitude alone helps or hurts tracking is not resolved: +0.006 ± 0.030 at 512 seeds, and no direction is claimed. |
| H3 | Protective sensing costs more per unit of region and less in total | Falls, in the useful direction. Cheaper on both counts, by about 8.5x per region point. |
| H4 | Sampling triggered by the agent's own intervention beats a clock | Falls outright. Scored against a no-sensor control the trigger is inert, holding to within 0.002 of what drawing zero samples holds at every disturbance level. |
7. A result of ours that was wrong, and how it was caught
Three. Two were errors of method and the third was an error of numerics.
The first version of this bench reported the attitude-alone tracking gap as 1.078 against 0.973. Per-episode spread of task error is about 0.25, so sixteen seeds resolve nothing finer than 0.06 while the effect is 0.03, and roughly a third of sixteen-seed schedules invert its sign. The direction survives at 512 seeds; the magnitude was about three times too large and is withdrawn. What moves it: the standard error is now printed beside every mean, so the size of an effect and the resolution of the sample are visible in the same glance.
The second reported a conditional law: efference-triggered sampling loses under disturbance and wins in a quiet world on a twentieth of the clock's samples. It does. So does drawing no samples at all, on a body that needs none. The comparison lacked a no-sensor arm, and the harness additionally granted the estimator a free exact state read on every barrier exit, which kept the trigger alive enough to resemble a policy. With the control restored and the read removed, the trigger is inert. What moves it: every sensing policy is now scored against the no-sensor arm, which turns "cheaper than the clock" into "worth more than nothing".
The third is the second morphology's stiffness sweep in §5, and it was the integrator rather than the harness. Forward Euler on a spring-damper is stable only while the step obeys DT·ω < 2ζ, and substituting ζ = C/(2√(kp·I)) and ω = √(kp/I) the inertia cancels and the whole condition becomes kp·DT < C: a bound on stiffness, which is the quantity being swept, and which no inspection of the plant makes visible. That capped the two-link body at 45 N·m/rad and the sweep ran to 138. Above the bound the integrator injects energy on every step and the divergence reads as physics: the region was published as rising to 0.678 at the threshold and collapsing to 0.000 by twice it, and a conclusion about morphology was drawn from that shape. The guard in place at the time bounded the sweep by DT·ω < 1.4, which is the criterion for a critically damped mode; this family is lightly damped, so that guard reported a ceiling of 344 where the true one was 45. Substepping the physics while holding the controller at 50 Hz removes the collapse entirely: 0.241 becomes 0.984 at 1.5 times the threshold and 0.000 becomes 1.000 at twice it, converged against a 32-substep control. There is no optimal compliance and no disagreement between the morphologies. The cart-pole was 1.33 times over its own ceiling at the top of its sweep, where the error was small enough to hide, and the published instrument let a reader drag the slider there. What moves it: each bench now derives its ceiling from its own damping constant and refuses a sweep that reaches it, and a preflight gate re-integrates every plant at four times the substep count and fails if a published region moves. The gate's self-test restores the original substep count and requires the defect to reappear.
What survives the first two is the same sentence, on firmer ground: the efference copy tells a machine what to subtract, not when to look. In this model the agent predicts its own contribution exactly, so its own actions are never the dominant source of prediction error, and there is nothing for an effort-triggered sampler to track.
8. What is this report careful not to claim?
Actuation energy here is physical, integrated as torque against velocity with a copper-loss term along the trajectory each policy actually produced. Sensing and compute energy are parameterised, calibrated so that a full always-on suite lands within an order of magnitude of actuation, which is the regime where the allocation question is interesting. Absolute joules are indicative and every result is a comparison at fixed calibration.
The body is a cart-pole with a lumped pivot spring-damper standing in for intrinsic actuator impedance. Whether the ratios here transfer to a legged or multi-contact body is not established by this report and should not be assumed in either direction. The region reported is a rescaled barrier-exit count over seeded rollouts, not a proof of set invariance; where this report says region it means that quantity, and the viability denominator this Institute reports against remains joules per viability-second.
9. What would sharpen this?
A body where the compute rail is separable on a shunt. Everything above is simulation with one physical term. The single measurement that would convert it is a small compliant body whose compute draws through its own instrumented rail, so that the three terms can be read apart rather than modelled apart. TR-2026-36 establishes the condition; this report supplies the reason to spend it on a sensing question.
A body that makes and breaks contact. The second morphology has been run and is §5; it agreed on the threshold and on the shape of the sweep, and inverted on what sensing buys. Both bodies have smooth dynamics, and a foot that lands is a different regime. Whether the sensor budget falls the same way when the barrier is a contact rather than an angle is the next sweep.
A per-read cost taken from datasheets rather than chosen. The four channel costs used here are calibrated, not measured. Real transducer and conversion figures would make the joules-per-region ranking a statement about hardware rather than about a parameterisation.
10. Conclusions
Sensing is the term of the task-energy sum that nobody separates, and it is the one term whose size the designer directly chooses. On the body measured here it is also the term whose necessary size is fixed by mechanism: a compliant pivot moves the region a blind controller holds from nothing to four fifths, and the sensing suite is left buying the task rather than the survival. The threshold that governs it is a property of the body's own equations and sits 7.3 percent above the textbook figure.
Read beside TR-2026-43, which sweeps the same body property against demonstration volume, the result generalises past sensing: compliance buys +0.742 of certified region there while 256 times more data buys −0.027. A dataset is a recording of sensor streams, so what never crosses a transducer is missing from both, and the two budgets share a floor set in the mechanism.
Two of the four hypotheses stated here were wrong in ways a control arm and a standard error would have caught. A third correction was not a hypothesis at all: an integrator that could not represent the stiffnesses being swept, which turned a monotone curve into a peak and a collapse and very nearly turned that into a claim about morphology. All three are withdrawn in §7 with the change that prevents a recurrence. The methods that produced those corrections cost less than the corrections did.
11. The forcing function
| What is missing | Why | What builds it | What becomes possible |
|---|---|---|---|
| Sensing separated on a real meter | The established split is compute against non-compute; sensing sits inside the second half undifferentiated | A compliant body with the compute and sensor rails instrumented separately | A sensor suite chosen by joules per unit of viability rather than by capability, before the perception stack is written |
| A third morphology, with contact | Two bodies gave two different answers about which vector sensing serves, and both have smooth dynamics | The same sweep on a body that makes and breaks contact, where the barrier is a foot rather than an angle | A bill of materials in which mechanical compliance and sensor count are traded explicitly, on the class of body Physical AI actually ships |
| Per-read costs from datasheets | The channel costs here are calibrated rather than measured | Transducer and conversion figures for a real sensor set | A ranking that is a statement about available hardware |
| The shared floor stated as a design rule | The sensing bill and the data bill are treated as separate budgets to trade against each other, when both are bounded below by the same mechanical property | Sweeping compliance against a sensor set and a demonstration count on one body, which TR-2026-42 §3 and TR-2026-43 §5 now do separately | A programme that stops answering a thin sensor budget by collecting more demonstrations, because it can see that both buy the same missing information |
| A course that teaches the loop's energy | Energy is taught as a datacentre or trajectory-planning subject; embodiment as perception and policy | PAI-330, twelve labs, benches that can fall | Engineers who price a sense before specifying it, and who write a disclosure that refuses what their setup cannot support |
References
- Institute for Physical AI @ JBI, Energy Observability in Embodied Systems, Technical Report TR-2026-36. companion report, read in full
- Institute for Physical AI @ JBI, The Physical AI Hardware Lottery, Technical Report TR-2026-40, §10, three denominators and the non-compute bound. companion report, read in full
- Institute for Physical AI @ JBI, Sim to Real: The Road to Physical Agency, Technical Report TR-2026-35, twenty-three regions in eighteen languages. companion report, read in full
- Institute for Physical AI @ JBI, The Two Operating Systems: What a Body Carries That No Dataset Contains, Technical Report TR-2026-43, §5, the substrate-against-data sweep on this plant. companion report, read in full
- ISO/TS 25213, Robotics: Test methods for measuring the energy consumption of robots: 6-Axis articulated industrial robots, ISO/TC 299, ICS 25.040.30, stage 60.00, effective 26 June 2026. catalogue record and scope verified; clause text not read
- E. von Holst and H. Mittelstaedt, "Das Reafferenzprinzip: Wechselwirkungen zwischen Zentralnervensystem und Peripherie," Die Naturwissenschaften 37, 464–476, 1950. read in translation
- W. Zhu et al., "Identifying Important Sensory Feedback for Learning Locomotion Skills," arXiv:2306.17101. read in full
- D. Attwell and S. B. Laughlin, "An Energy Budget for Signaling in the Grey Matter of the Brain," J. Cereb. Blood Flow Metab. 21, 1133–1145, 2001. read in full
- Institute for Physical AI @ JBI, Charlot Lab, the SENSE/1 sensor-ledger bench, the SENSE/2 second-morphology bench and the FOCUS/1 allocation bench, 512 and 16 seeds respectively, dispersion reported on every mean, versioned at
datasets/loop-energy/bench/. this report's instruments; every figure reproduced from them