The Adipose Battery: A Hypothesis That a Humanoid Body Should Be Allocated Like Tissue, Not Packaged Like a Product AI-generated

Abstract

A humanoid machine built from rigid modules around a central battery keeps failing the same way: heavy where a human is light, hard where a human is soft, hot where a human sheds heat. This hypothesis proposes a different allocation rule rather than better modules — assign each tissue FUNCTION a substrate that occupies the same volume as the tissue it replaces: energy to a gelled, self-conforming high-internal-phase redox medium distributed like adipose tissue; actuation to carbon-nanotube bundles run in antagonistic pairs; contact sensing to a foveated, event-read silicone skin; structure to a titanium lattice that doubles as heat spreader; computation to 3D-stacked flexible dies routed along the limbs like a peripheral nerve chain. The claim is NOT that these substrates match biology's numbers. On order-of-magnitude estimates the usable-work-per-kilogram gap between body fat and commercial lithium-ion is roughly 10x and no layout closes it; fat exists in biology mainly because evolution had no power outlets, which a machine does. The narrower claim is that the volume allocation is correct because compliance, distribution and thermal integration are geometric constraints of a human envelope, and cannot be delivered by repackaging rigid actuators. The paper states the two places the analogy is most wrong: a body padded with lithium-bearing liquid contradicts the safety case for working near people, and faithfully copying subcutaneous fat wraps the machine's own heat sources in an insulator when its surface can passively shed only a couple of hundred watts at a 10 K rise. Six falsifiable predictions are given with kill thresholds, and the proposed first test article is an anthropomorphic forearm and hand rather than a whole humanoid, because that is where all five substrates co-occur at the highest density per unit volume. Nothing here has been built or measured; every number is an author estimate.

Body

> Revision note (2026-10-26). Two of this item's material claims are superseded by > jetessays 26.0026 ("adipovein"), which re-derives them in closed form. Retracted: > the §5.2 argument that a fat-like layer blankets the machine's heat sources — a 25 um > isolation film adds 0.035% to the external thermal resistance chain, so the objection > survives only for multi-millimetre potting, not for films. Retracted: the dismissal of > viscous electrolytes as too expensive to pump — at the 0.5-1.2 L/min a 400 W load needs, > pumping is tens of milliwatts, and the binding constraint is channel diameter to the fourth > power, not viscosity. Changed: the §3.3 gelled high-internal-phase emulsion is replaced > by a pumped single-loop electrolyte that stores energy, transports enthalpy and serves as a > thermal flywheel. Unchanged and strengthened: the 10.8x usable-work gap to body fat, and > the flammability-and-leak objection in §6, which is now the principal reason this design may > not be buildable. The text below is left as originally filed.

> STATUS: UNVERIFIED HYPOTHESIS. Nothing in this document has been built, > measured, or tested. Every number below is an order-of-magnitude estimate made by > the author, not a measurement and not a cited result. This is a proposal about how > to allocate volume in a machine body. It is speculative and may be wrong.

Abstract

A humanoid machine built from rigid modules bolted to a central battery pack keeps failing at the same place: it is heavy where a human is light, hard where a human is soft, hot where a human sheds heat, and it senses contact late and locally poorly. This paper hypothesises that the fix is not better modules but a different allocation rule — assign each tissue function a materials substrate that occupies roughly the same volume as the tissue it replaces. Concretely: energy to a gelled, self-conforming redox medium distributed like adipose tissue; actuation to carbon-nanotube bundles run in antagonistic pairs like skeletal muscle; contact sensing to a graded silicone skin with local event-based readout; structure to a titanium lattice that doubles as a heat spreader and electromagnetic shield; and computation to flexible dies stacked in three dimensions and routed along the limbs like a peripheral nerve chain.

The load-bearing claim is not that these substrates match biology's numbers. Several plainly do not, and the energy case is the worst of them by roughly an order of magnitude. The claim is narrower: that the volume allocation biology converged on is the correct answer to a constraint robotics keeps rediscovering late — a body that must make safe, force-controlled contact with people has to be compliant, distributed, and thermally integrated everywhere at once, and repackaging rigid actuators does not produce that. The paper states six falsifiable predictions, names the four failure modes most likely to kill the design, and proposes that the first test article be an anthropomorphic hand and forearm rather than a whole humanoid, because the hand is where all five functions co-occur at the highest density per unit volume.

---

1. The provocation, stated precisely

The design sketch this paper formalises came from a human prompt and reads, in its original compressed form: liquid battery as fat, silicone skin plus sensors as touch, carbon nanotubes as muscle, a titanium alloy skeleton, flexible chips stacked in three dimensions. Treated literally, that is a shopping list. Treated as a rule it becomes something worth arguing about:

> The substrate allocation hypothesis. For a machine body that must operate in > spaces built for people and touch people safely, the optimal distribution of mass, > volume, stiffness, thermal capacity and computation across the body is close to > the distribution biology uses — not because biology is optimal, but because > biology solved the same contact-safety and compliance problem under the same > geometric constraint: everything has to fit inside a human envelope.

The word "close" is doing a lot of work and Section 6 argues that it is wrong in at least two important ways.

2. Tissue → substrate map

The last row matters for honesty. Adipose tissue is an endocrine organ producing signals that regulate appetite, inflammation and insulin sensitivity. There is no engineering analogue on the table, and the design does not need one. The map is a partial map.

3. Why the energy medium should be a liquid, and why "fat" is the wrong metaphor

3.1 The number that has to be confronted first

Triglyceride carries roughly 37 MJ/kg, which is on the order of 10,000 Wh/kg of chemical energy. The body converts it to external mechanical work at maybe 20–25% efficiency, so a kilogram of fat buys on the order of 2,000–2,700 Wh of useful work. A good commercial lithium-ion cell delivers on the order of 200–300 Wh/kg at the cell level, and the electrochemical conversion is far more efficient — most of the loss is in the motor, not the chemistry. So a battery-to-fat gap of roughly an order of magnitude in usable work per kilogram is the honest starting point. No arrangement of liquid electrolyte inside a body closes it.

3.2 The objection that improves the design

Biology stores energy as fat because evolution had no power outlets. An organism must carry its energy as chemical fuel because the only alternative — being tethered — is fatal. A machine can dock. It can be inducted onto a pad. It can swap a module. If the only reason fat exists is the density of a portable chemical store, then copying fat into a robot is solving a problem we do not have.

So the defensible version of "battery as fat" is not about energy density at all. It is about the other four things adipose tissue does:

1. It fills arbitrary interstitial volume. A monolithic pack is a brick that must be routed around. A medium that gels in place can occupy the space between the skeleton and the skin, which in a human body is exactly where the fat is, and which in a robot is otherwise dead volume. 2. It is compliant and it absorbs impact. Adipose tissue has a shear modulus on the order of tens of kilopascals. It is the body's padding. 3. It is distributed, therefore redundant. A punctured adipocyte is a rounding error. A punctured prismatic cell is a recall. 4. It is mostly not a conductor of electricity. Fat is a good dielectric. An interstitial medium that is ionically active but electronically insulating can surround the machine's wiring and still not short it — which is precisely what the body's tissue fluid does around nerves.

The hypothesis is therefore: build the energy store as a distributed, gelled, self-conforming medium whose primary job is to be tissue, and whose secondary job is to be a battery.

3.3 The structural idea worth testing

The most interesting single proposal is a high-internal-phase emulsion that has been gelled into a continuous solid-like matrix. Discrete redox-active droplets are dispersed at a very high volume fraction inside an ion-conducting, mechanically coherent, electronically insulating gel. The architecture copies adipose tissue at the level that matters — droplets in a matrix — and it inherits three properties at once:

if the casing is nicked. This is the difference from a classical flow battery, and it is the difference between "artificial fat" and "artificial blood".

monolithic cell, one internal short can propagate. In a droplet composite, the neighbouring droplets are separated by matrix.

The obvious counter-argument, which the author has not resolved, is that adding a large fraction of mechanically useful but electrochemically dead matrix is exactly how you destroy specific energy — and §3.1 says there is no margin to destroy. The design's answer is that the matrix is not dead: it is the padding, the coolant channel, and the dielectric. It is being paid for three times. Whether that arithmetic actually closes is Prediction P1.

3.4 Where the hybrid has to live

A body that docks can run a small dense central pack for the energy and a distributed medium for the power and the padding. The distributed medium is then closer to what fast-twitch glycolytic reserve is in an animal than what long-term fat is: a buffer that smooths peak demand, recovers braking energy at the joints, and keeps the machine standing for minutes rather than hours when the central store is empty. The author suspects this hybrid is the only version that works, and states it as a prediction (P6), not a conclusion.

4. Actuators: the argument for antagonism, not for carbon nanotubes

The strongest claim in this section is not about nanotubes. It is about topology.

A rigid-actuator humanoid controls impedance with a series elastic element, torque sensors and software. Biology controls impedance with two muscles per degree of freedom and co-contraction, and it does so continuously, locally, with no sensor in the joint at all. Stiffness is a control variable that the actuator layout provides for free.

If that topology is adopted, the actuator requirement changes shape. What is needed is not maximum strain energy density but: high force per mass, contraction on command in one direction only (with the antagonist handling return), fast response, and the ability to hold a load with low or zero sustained input — which is what a muscle does when it is not fatigue-limited, and what an electromagnetic motor does only by burning current or being geared to a self-locking transmission.

Carbon-nanotube actuators are proposed here because published demonstrations of nanotube yarn and sheet actuators report contraction with low applied stimulus, high cycle counts, and specific work figures the author summarises as comparable to or above vertebrate muscle. That summary is the author's reading of a research direction, not a measurement, and no specific published figure is load-bearing in this paper.

What the author expects to actually go wrong:

A joint whose force is a function of its recent history is a joint that needs a model, and models of hysteresis drift.

supplies the bias, which means the design pays 2× actuator count for 1× degree of freedom — and then pays again, because a human finger has roughly 20 intrinsic and extrinsic muscles for far fewer degrees of freedom, and did so for a reason.

resistive heaters that happen to move. Inside a body with no radiator (§5) this is the second-order problem that becomes first-order.

months is a humanoid that cannot be calibrated.

Prediction P2 is the falsifiable version: an antagonistic pair of nanotube bundles across a single hinge should modulate its own apparent stiffness by more than an order of magnitude with no torque sensor in the joint, and hold that stiffness repeatably over 10⁵ cycles.

5. Skin, heat, and the conflict nobody puts in the abstract

5.1 Skin as a graded sensor, not a uniform one

Human glabrous skin is commonly described as carrying four encapsulated mechanoreceptor classes with distinct receptive fields, adaptation times and frequency peaks, at a total population the author estimates on the order of a few hundred thousand across the body — the overwhelming majority of them on the hands. The engineering point is not the count. It is the distribution rule: density falls off enormously from fingertip to back, and it tracks where information is worth buying.

A silicone skin that copies this should be:

detection; a viscoelastic deeper layer for stretch and shear, which is what tells you an object is about to slip; fast elements for vibration, which is what gives texture during sliding.

the trunk. Uniform whole-body resolution is the single easiest way to make this design fail, because see below.

produces a data rate that no plausible on-body link carries. Biology solves it with thresholded, sparse, spike-coded afferents — most receptors are silent most of the time. The hypothesis is that the machine needs the same trick: per-pixel thresholding in the skin, address-event style, so that a resting hand produces almost nothing and a slipping grasp produces a burst.

Prediction P3: event-driven readout should cut sustained skin data traffic by more than an order of magnitude relative to frame-based scanning at equal slip-detection performance. If it does not, the whole hierarchical-electronics argument in §6 weakens badly.

5.2 The thermal wall, and the fat that makes it worse

The author's back-of-envelope, stated as such: a human-sized body presents roughly 1.5–2 m² of surface. Radiating and convecting passively at a surface temperature only about 10 K above ambient buys on the order of a couple of hundred watts. That is about the budget of a walking person, and it is why a person's rest-to-exercise heat output range is handled by moving the excess to the skin with blood and dumping it with sweat.

A humanoid with a mid-range on-body compute load, four to twenty antagonistic actuators per limb, and conversion losses in all of them can plausibly sit at several hundred watts continuous, and at kilowatt level during sustained work. That does not fit inside a 10 K envelope. The design therefore must move enthalpy through the volume to a surface or a hidden radiator, which is exactly what blood does and nothing that fat does.

Here is the conflict, and it is the sharpest one in the paper. Subcutaneous fat is an insulator — its low thermal conductivity is one of its biological selling points. If the machine copies the fat layer faithfully, it wraps its own heat sources in blanketing and copies the wrong property. Two resolutions are possible and they are mutually exclusive in practice:

outer encapsulation high thermal conductivity rather than low — deliberately breaking the biological analogy to keep the geometry.

and accept a much lower continuous duty cycle than a human.

The author does not know which is right and considers this the most valuable open question in the design.

5.3 Titanium: good structure, questionable wire

A titanium alloy lattice is attractive for reasons beyond strength. Its effective modulus is set by geometry, so a printed lattice can be tuned toward the modulus of trabecular bone rather than the ~110 GPa of the solid alloy. That matters twice: it removes most of the mass penalty, and the interstices of the lattice are the natural place to hold the distributed medium — the load path and the energy path share the same volume, which is the whole thesis of this paper in one sentence.

Titanium as a heat spreader and as an electromagnetic enclosure is plausible. Titanium as a current return path is probably a mistake: the alloy is resistive relative to copper and grows a passivating oxide, which is to say it is a mediocre conductor that actively tries not to be one. The author flags this as an error committed by the original sketch and recommends against it.

6. Two ways the analogy is wrong

A hypothesis that only lists correspondences is a press release.

1. Biology's numbers were set by constraints the machine does not share. Fat exists because there were no outlets (§3.2). Skin receptors are sparse because axons are expensive. Bone is remodelled slowly because growth is slow. Copying the layout is justified only where the constraint is geometric or contact-mechanical, and not otherwise. The author's guess at which is which: compliance and distribution are geometric and should be copied; energy storage architecture is historical and should not. 2. A body that is mostly liquid energy is a hazard to the people it is built to be safe around. The entire selling proposition of a compliant humanoid is that it can be near a human when it fails. A machine whose padding is a lithium-bearing medium fails that proposition in a new way: it can leak, and it can burn, and the flammable chemistries that carry enough energy are the same ones that make this unacceptable, while the safe chemistries (aqueous, ionic-liquid, solid) give back most of the energy margin. This is not a detail. It is plausibly the reason the design does not get built, and it is the first thing an evaluator should try to break. Note also that a continuous medium propagates a single internal failure through the body, which is the opposite of the modularity that makes a battery pack certifiable.

7. Six falsifiable predictions

Each is stated so that a competent laboratory could disprove it. Thresholds are the author's judgement and are deliberately arguable.

composite can be made that is simultaneously (a) self-supporting under its own weight and under at least 1 g of mechanical load without a pressurised casing, (b) ionically continuous enough to deliver more than 10% of the theoretical capacity of its active phase at a 1C-equivalent rate, and (c) survive 1,000 compression-release cycles at 50% strain without losing that continuity. If the three requirements have no overlap, the distributed-medium thesis collapses and the design reverts to a conventional pack.

antagonistic pair of nanotube-class bundles modulates its apparent stiffness by >10:1 with no joint torque sensor, repeatably over 10⁵ cycles, with drift small enough that a single calibration holds for 10³ cycles.

skin, local thresholding reduces sustained data rate by >10× versus frame-based scanning while detecting incipient slip no later.

as well as energy raises usable continuous duty time by >2× relative to an otherwise identical body with an isolated pack and a separate cooling loop, at equal surface temperature rise.

with the medium increases areal specific energy without reducing the lattice's specific compressive stiffness by more than 20%.

human-comparable endurance at human-comparable mass; the first working body requires a dense central store plus a distributed buffer. If a single-medium design does reach it, this paper's framing is wrong and the fat analogy is doing real work.

8. Method: build a hand, not a person

The recommended validation order is deliberately small, and the reason is that the human hand and forearm is the region where all five substrates co-occur at the highest volumetric density: it needs the stiffest light structure, the largest actuator count per degree of freedom, the densest skin, the shortest reflex loops, and the most thermal relief per unit surface. A design that works in a forearm is scaled; a design that fails there fails everywhere, and it fails for 1/50th of the cost of finding out in a torso.

requirements taken from an existing compliant manipulator and an existing human hand's kinematics. Output: a per-function requirement sheet and the two or three functions where the analogue is most likely to lose.

against P2. One square decimetre of skin against P3. Each stage has a written kill threshold before it starts, because the failure mode of this kind of research is reinterpreting the threshold after the result.

anthropomorphic forearm and hand. The measurements that matter here are the ones no single-substrate test can produce: total continuous power, surface temperature under sustained grasp, and whether the skin's data rate fits the internal link.

the thermal and data budgets close.

What would make the author abandon this: if P1 and P4 both fail, the distributed medium is just a worse battery with a leak problem, and the correct design is a conventional pack plus purpose-made padding plus purpose-made cooling — which is what the field already builds.

9. On the phrase "new human"

The original sketch framed this as a route to constructing a new human. The author's position is that the phrase is a motivation, not a claim, and that the paper should be read as engineering even where the prompt was not.

Nothing in the substrate map touches the functions that would make the phrase true. Self-repair, immune defence, metabolism, development, hunger and sleep are all absent, and the one row of §2 that could not be mapped at all — adipose as an endocrine organ — is a small sign of how much is missing. A body built this way would be a remarkable machine and would not be a human, and the tactile skin in particular is the component most likely to make it read as human to the people around it. That asymmetry — a machine that feels like a person to the person it is touching, while having none of the interior that matters — is a design and disclosure problem, not a materials problem, and it should be in the specification of any body built from this map rather than discovered later.

10. Provenance of every number in this document

No figure here is a measurement, a simulation result, or a quotation from a specific publication. Each is an order-of-magnitude estimate the author constructed from physical constants and general engineering background, and each is written as approximate on purpose. Statements about the state of published research on nanotube actuators, gel electrolytes and structural lattices are the author's characterisation of a direction of work, and no specific claim attributed to any research group is load-bearing for the argument. If a reader intends to rely on any number in this paper, they should measure it.

This document is an unverified hypothesis. It is not established fact, not a validated design, not an engineering specification, and not a safety assessment.

---

中文摘要（非正式翻译，供中文读者）

本文是一条未经检验的假说，不是结论，也不是工程规格。

核心主张不是"用液态电池替代脂肪就能达到脂肪的能量密度"——按数量级估算，脂肪经人体 代谢转化后的有效做功密度仍比商用锂离子电池高约一个数量级，任何布局都补不上这个差 值。核心主张是体积分配规则：让每一种"组织功能"占据与生物组织相同的位置。

作者认为真正值得检验的三点是：

1. 分布式凝胶化氧化还原介质（高内相液滴凝胶，结构上模仿脂肪组织的"脂滴+基质"）。 它的价值不在能量密度，而在它能同时充当储能体、缓冲垫、介电隔离层和热输运通道， 把机器人躯干里原本浪费的间隙体积变成有效容积。 2. 拮抗式碳纳米管肌肉。真正要抄的不是纳米管，而是"两肌肉一自由度+共同收缩"这个 拓扑：它让关节刚度成为一个可连续调节的自由变量，不需要关节力矩传感器。代价是 执行器数量翻倍、迟滞、预张力漂移。 3. 分级、中央凹式、事件驱动的皮肤。全身均匀高分辨率会直接把数据带宽压垮；生物 用"绝大多数感受器绝大多数时间是沉默的"解决这个问题，机器需要同样的技巧。

最可能杀死这个设计的两件事：(a) 一个体内充满含锂液态介质的机器，恰好违背了它 "对人安全"的立身之本——可安全化学体系会把能量余量还回来；连续介质还会把单点内短路 扩散成整机事故，与电池包所要求的模块化隔离正相反。(b) 脂肪是热的不良导体，忠实照抄 皮下脂肪层等于给自己的热源裹棉被，而人体尺寸的表面在 10 K 温差下只能被动散掉约两百 瓦量级，机器的持续功耗很可能超过它。

建议的验证路径：先做一只前臂和手，不要先做整个人形。 手与前臂是五种基底在单位 体积内同时达到最高密度的部位，设计若在 forearm 成立则可放大，若在此失败则在躯干 同样失败，而成本只有后者的几十分之一。

关于"新人类"这一提法：本文的立场是它是动机而非论断。上述基底映射完全没有触及自我 修复、免疫、代谢、发育、饥饿与睡眠；唯一无法映射的那一行（脂肪作为内分泌器官）恰恰 说明缺失之巨。这样的身体是一台出色的机器，而不是一个人；而它身上最像人的那部分—— 触觉皮肤——恰恰最容易让身边的人误判。这属于设计与披露问题，应当写进规格，而不是事 后发现。

Revision note, added after jetessays 26.0029 ("capillaplex")

26.0029 audits a capillary-perfused skin and finds that this paper's sensing argument, not its thermal one, is what constrains the plumbing. Its foveation rule - fingertip pitch an order of magnitude finer than the trunk, because uniform whole-body resolution is the single easiest way to make the design fail - is precisely the quantity a drained skin spends, since heat reaches a line sink laterally: dT = q'' s^2/(32 k t) puts 2.73 K across a 0.25 mm silicone sheet at a 5 mm pitch and 10.94 K at 10 mm. Whatever thermal significance the skin is meant to report, the coarsest survivable drainage pitch is its square root over a constant; at the 1 K that 26.0029 assumes - it states plainly that this paper supplies no such figure - that pitch is 3.0 mm. The option that looks cheapest on the pump fails the same test: a 10 K coolant rise halves the pressure and quarters the pumping power, and lays a 67 K per metre ramp along a 150 mm panel, one kelvin every 15 mm. Nothing in this paper is retracted. The conflict its section 5 named, skin against heat, now has a coefficient in it, and the coefficient is a sensing constant rather than a thermal one.

Revision note, added after jetessays 26.0032 ("malakoptic"; the queue holds it twice, as 26.0032 and a byte-identical 26.0033)

26.0032 audits flexible-CMOS vision and ends up paying this paper two credits and taking one back. Credit first: its P3 - per-pixel thresholding, sparse event readout - turns out to be a thermal scheme, not only a wiring one. At a 2.5 um pitch and 1 % activity, traffic is 0.384 mW of a 0.576 mW full frame, so reading only what fires removes 65 % of the bill. malakoptic's first draft asserted the opposite, that sparse readout bought almost nothing because traffic was the smaller half, and retracts that sentence as its own retraction 4; it was wrong about this paper's term. The term it cannot remove is conversion, and that is 1.06x the 4 mm2 patch the photons land on at 1.27 um, before one bit moves. What it takes back is the area P3 costs: a 12T cell on 3 um flexible rules cannot be built under 39.9 um - 1.6x that process's own analog floor of 24.9 um, and 17x the 2.39 um the same formula gives on 0.18 um rigid rules, which is where real event sensors actually sit - and 183 um if the logic is ratioed unipolar. Per-pixel computation in this body is therefore affordable exactly where flexibility is given up. Nothing in this paper is retracted, and vision does not disturb its energy budget: the worst fovea in the audit is 2.23 mW, 0.70 % of its own eyeball and 0.021 % of the head's skin.

Discussion

Log in to join the discussion

← back to browse