Everything the Institute has published, in one place.

What a Decision Costs

A typed-decision model answers without generating text, and the saving is real. It is also bounded by a quantity the claims do not mention: the accuracy of whatever decides which decisions go to it.

A class of models shipped in 2026 that answers a question without writing a sentence: unstructured state in, a typed value and a calibrated probability out, in one forward pass with no autoregressive loop. The case for it is an energy case, and vendors put the saving at a hundredfold and above. This report accepts the mechanism and computes what bounds it. A decision routed to the typed engine and answered wrongly pa…

Read online →↓ PDF · 0.4 MBResearch / Position · Preprint v1.0

What Do You Type?

The software toolchain of thermodynamic computing, surveyed across five language ecosystems. The hardware argument is settled enough to fund. The interface question, what a person or an agent actually writes and what comes back, is not, and it is now the binding constraint.

Thermodynamic and Ising-style computers are past the point where the hardware argument decides adoption. What decides it now is the interface: what a user writes, and what the machine writes back. We surveyed the toolchain layer in five language ecosystems, using Japanese, Chinese, English, German and Korean sources, and found four distinct interface designs, each coherent and each solving a different problem: a mode…

Read online →↓ PDF · 0.5 MBSurvey / Position · Preprint v1.1

Where Is the Energy Reporting?

Embodied energy is measured in the research literature and is being standardised for the industrial arm. It is reported nowhere a buyer, an operator or a regulator can act on. The binding constraint is disclosure, not instrumentation, and that makes it a qualification problem before it is an engineering one.

A humanoid that walks efficiently makes its computer the thing to argue about. Hold the compute module fixed at its published envelope and improve only the gait, from the best figure the earlier literature carried to one reported in 2026, and the computer's share of shift power roughly triples. Nothing about the brain changed; the term it competes with fell fourfold. That is a good problem, because it means the machi…

Read online →↓ PDF · 0.4 MBResearch / Position · Preprint v1.0

Six Ways a Body Cheats

The mechanisms human movement uses to make acrobatics affordable, five of them priced against an identified power model, and the four predictions of ours that the pricing falsified.

Humanoid robots now perform acrobatics, and they do it at a cost of transport roughly sixteen times a human's 3. The gap is usually read as a control problem. This report reads it as an energetics one and asks which of the mechanisms biology uses actually transfer. Six are surveyed and five are put on a bench sharing the loss model of TR-2026-41. The organising result is that a body and an electric motor pay for the …

Read online →↓ PDF · 0.4 MBResearch / Modelled Results · Preprint v1

The Unknown Fraction

Agricultural sensing reports what it measured. It does not report what it could not see. Three instruments over three unrelated physics find the same division, and four industries already pay for the quantity none of them computes.

The unknown fraction of a sensed volume is computable, and it equals the probability of missing whatever happens in it. That is a useful pair, because a quantity this review did not locate in any published sensing system turns out to be computable by anyone, from the geometry they already have. Agriculture is the hard case and therefore the interesting one: it is instrumented at two extremes, satellites resolving the…

Read online →↓ PDF · 0.4 MBResearch / Position · Preprint v1.0

The Two Operating Systems

A natural agent runs a substrate that carries most of the behaviour and an information system that rides on top of it. The data thesis for Physical AI scales only the second, and three independent mathematics say why that is not enough.

Companies and academic centres are converging on the view that a sufficient volume of data will unlock Physical AI, and the strongest evidence is real: a log-linear scaling law between egocentric human video and robot policy performance 3. This report assembles the case that the thesis mistakes what a natural agent is. A body runs two systems: a substrate of compliance, preflexes, pattern generators and regulation th…

Read online →↓ PDF · 0.4 MBResearch / Position + Measurement · Preprint v2

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.

Read online →↓ PDF · 0.4 MBResearch / Position · Preprint v1.0

Joules per Punch

Where the energy of a humanoid strike goes, why a segmented spine costs more than one waist joint at every demand both can meet, and why it is still the only body that throws the hardest strike.

In January 2026 two Unitree G1 humanoids boxed at CES, and the commentary settled on one diagnosis: the motion looks wrong because the control policy is weak 7. The kinematic diagnosis runs deeper. The classical force decomposition of the boxer's straight punch places 38.5 percent of its force in leg drive and 37.3 percent in trunk rotation, against 24.1 percent in the arm 1, and the base G1 carries one yaw joint whe…

Read online →↓ PDF · 0.6 MBResearch / Modelled Results · Preprint v1.4

The Physical AI Hardware Lottery

Energy-based models, post-von-Neumann architecture, and how a substrate shapes which ideas get developed.

In 2020 Sara Hooker described a mechanism she called the hardware lottery: a research idea advances partly because it suits the available hardware, and hardware can therefore delay a line of work by making it look unproductive 1. This report traces that mechanism through energy-based models, and the tracing is mechanical rather than interpretive. Fitting an energy-based model by maximum likelihood requires the partit…

Read online →↓ PDF · 0.7 MBResearch / Position · Preprint v1

The Portable Core

Physical AI and the logistics of nuclear energy. The world is building the autonomy an unattended plant needs. It is not building the layer that lets the plant prove what it did.

Every embodied system ends in a joule supply chain, and portable nuclear power is the first energy technology that ships to its load rather than tethering the load to a grid. The dependency runs both ways: a sub-20 MWe plant cannot carry a 24/7 licensed crew, so the staffing model that works at 1 GW is the cost floor that kills 1 MW, and physical AI becomes not the customer of portable nuclear but its crew. This repo…

Read online →↓ PDF · 0.4 MBResearch / Position · Preprint v1

Sim to Real: The Road to Physical Agency

What separates a policy that works in simulation from one that works on a body, read in twenty-three languages against the Institute's own gaps.

A policy that works in simulation and fails on a body has met one of four walls: the world model is wrong, the action-conditioned transition function is wrong, a sensing channel has degraded rather than failed, or the body itself has drifted. This review asks what the record measures at each wall, and it asks in twenty-three languages, because the English-language conversation is not the record. Four evidence bases a…

Read online →↓ PDF · 2.2 MBResearch / Review · Preprint v3

RELAX/1

Ten operations, an MLIR and NIR-style exchange form, and a lowering matrix across substrates.

If relaxation-style computing is to run on thermodynamic samplers, Ising machines, neuromorphic fabrics and ordinary GPUs alike, it needs a dialect that lowers to all of them. RELAX/1 specifies ten operations with a reference lowering, so that one program can be executed by several kinds of physics and the receipts compose across them.…

Read online →↓ PDF · 0.3 MBDraft v0.1 · CC0

OER/1 and DRIFT/1

A receipt schema for energy claims, and a benchmark protocol for learning under drift.

Alternative-computing waves before this one (analog VLSI, optical computing, memristive crossbars, quantum annealing) each stalled at the same point on the trajectory: a claimed efficiency advantage no shared measurement protocol could confirm or refute. Each is still in progress, and each moved further once one was adopted. Naming the binding constraint as measurement methodology rather than physics is the premise o…

Read online →↓ PDF · 0.4 MBDraft v0.1 · CC0

A Certificate That Bounds Danger Does Not Bound Waste

A safety certificate constrains where a machine may go, and leaves free what it may spend getting there. That freedom is a measurable channel, and the same measurement that would finally make robot energy comparable is the one that closes it.

Certificate-gated control is a strong answer to a compromised policy: a backdoor grants an adversary an arbitrary policy, but not an arbitrary trajectory, because every action must still pass a check made on the device. This report asks what that leaves open, and answers with a number. A certificate distinguishes safe from unsafe actions; it does not distinguish cheap ones from expensive ones. On a measured autonomou…

Read online →↓ PDF · 0.4 MBResearch / Position · Preprint v1

Can a Machine Prove It Is Itself?

A drivetrain carries a measurable manufacturing signature. This report prices it in bits, and the price decides what it can be used for.

A body that can be stolen, altered or substituted needs a way to demonstrate it is the body that was certified, and software cannot supply one because software is what an adversary replaces. Manufacturing variance is the classical answer: silicon physically unclonable functions turn fabrication accidents into identity. This report asks whether a drivetrain does the same, and answers entirely from published coefficien…

Read online →↓ PDF · 0.5 MBResearch / Position · Preprint v1

Energy Observability in Embodied Systems

What the meter can prove about a machine, and where. Morphology decides whether computation shows up in the joules at all, and on the bodies Physical AI actually uses this review did not locate a method that settles it, which makes the instrument itself the opportunity.

Software asserts; energy is spent. A power trace is the one account of a machine's activity that its own code cannot simply write, which makes energy a candidate substrate for a machine that must demonstrate its behaviour rather than claim it. This report asks how far that demonstration reaches on an embodied system, and answers from measurement rather than from architecture. Three findings carry it. First, the compu…

Read online →↓ PDF · 0.5 MBResearch / Position · Preprint v1.1

Physical AI and the Department of Transportation

Opportunities for automation across diverse geographies: what the work is, which constraint binds it, and how far along each class actually sits.

A transportation estate is not one estate. This review partitions the United States road network into four terrain and climate classes, anchors each in a named counterparty, and asks of each the same three questions: what is the work, which constraint binds it, and where on the trajectory does it sit today. The national baseline is 8,865,888.676 lane-miles across 4,208,454 centerline miles 1,2, of which state highway…

Read online →↓ PDF · 0.8 MBResearch / Review · Preprint v1

Economic and Environmental Impact of Distributed AI for Physical AI

Centralised hyperscale facilities are the default location for the computation that serves physical AI. This review surveys the alternatives, grades the measured evidence for each, tests eleven stated hypotheses against it, and computes the physical limits that bound the most distributed end of the range.

Physical AI raises a siting question. The computation that serves a machine moving through the world can be carried by the machine, placed in a nearby facility, embedded in a network operator's infrastructure, or concentrated in a centralised hyperscale campus, and the choice carries economic and environmental consequences that are argued about more often than they are measured. This review states eleven hypotheses, …

Read online →↓ PDF · 1.0 MBResearch / Review · Preprint v1

Physical AI and Logistics Opportunities in Agriculture

US agriculture is counted well at the top and thinly where a machine has to work. This review sizes what Physical AI can take on across the estate, the labour market and the logistics layer where the sources allow and records where they do not, grades every figure it locates, tests eleven numbered hypotheses, and places each application on the trajectory with the constraint that currently binds it and the measured change that would move it.

This review is organised as eleven numbered hypotheses, H1 to H11. Each is stated so that it can be checked against a named source carrying a verification grade, and each resolves to a position on the trajectory of agricultural automation, with the constraint that currently binds that position named and a quantity attached to the change that would move it. The 2022 Census of Agriculture counted 1,900,487 farms on 880…

Read online →↓ PDF · 1.2 MBResearch / Review · Preprint v1

The Future of Commerce Using Physical AI

Eleven hypotheses about physical AI in commerce, tested against eighteen graded deployment records, three BLS series, four frontline experiments of which three are randomised, and the physical quantities that set what sensing and manipulation cost. Each resolves to a position on the trajectory, a named binding constraint and the measured change that moves it.

Physical AI has been installed on the American selling floor at every scale from a single trial store to a chain-wide rollout, and this review reads that record as a set of trajectory positions rather than as a set of outcomes. Eleven hypotheses are stated, tested against eighteen graded deployment records, three Bureau of Labor Statistics employment series with the CPI-U used separately to deflate earnings, and four…

Read online →↓ PDF · 1.0 MBResearch / Review · Preprint v1

Physical AI and Logistics Management for Telecommunications

Telecom operators run one of the largest distributed physical estates in any industry. This review states ten hypotheses about where Physical AI is carrying value across that estate, tests each against graded evidence, and resolves each to a position on the trajectory and the measured change that would move it, nine to a binding constraint and the tenth to the measurement that would name one.

Telecommunications operators run one of the largest distributed physical estates in any industry and hold an approximate record of its contents. CTIA counts 447,605 US cell sites at the end of 2024 6. WIA's 2025 report, modelled by iGR, estimates 254,850 macrocells and 158,500 purpose-built towers 4, and this review did not locate a published methodology from either body that reconciles the two. The buried-plant reco…

Read online →↓ PDF · 0.8 MBResearch / Review · Preprint v1

Physical AI in Space and the Orbital Data Center Economy

Above the Kármán line there is no labour pool, which makes Physical AI the space workforce itself, at any price. This review prices that claim against the 2025–26 flight record, the launch-cost curve, and 83 sourced company and program records.

The space economy crossed $626 billion in 20251 and is projected to reach $1.0–1.8 trillion by the mid-2030s2,3. Two theses inside it are moving from paper to hardware: orbital data centers, solar-powered compute placed in orbit to escape terrestrial power, land and cooling limits, and Physical AI, artificial intelligence embodied in machines that sense, move and manipulate. This review argues they are one story. Abo…

Read online →↓ PDF · 0.9 MBResearch / Review · Preprint v0.1

Quantum Information Science for Edge Physical AI

A robot at the edge has a fixed battery and no datacenter, so the question is precise: where does quantum information science help a body compute on its own hardware today? The grounded answer is that most of the near-term value is quantum-inspired structure run on classical silicon you can already carry, and this paper measures exactly where it is real, where it is not, and why.

Quantum information science reaches embodied AI through several doors, and only some of them open at the edge today. This paper walks each with a measurement, not a promise. (1) Tensor-network compression (the mathematics of entanglement, borrowed from many-body physics) is the one that ships now: a control policy trained in factored form matches the full policy's accuracy at 32× fewer parameters, and a pure-Rust fac…

Read online →↓ PDF · 0.5 MBSurvey / Position · Preprint v0.1

Building the Energy Compute Future

Why performance claims for new machines keep collapsing, which claims survive the attack, and how to state the two of them together.

A new computing machine is announced with a large speedup. Within a few years an ordinary computer running a cleverer algorithm reaches the same place, and the advantage is gone. This has happened often enough to have a name, dequantization, and it is the central obstacle to arguing that any novel paradigm is worth building. This report separates two claims that are usually made in one breath. The first is that the n…

Read online →↓ PDF · 0.2 MBResearch / Method · Preprint v0.1

The Adaptive Unit

Why a machine that keeps learning needs a richer neuron, what dendrites buy, and whether the better primitive can win on the hardware we already have.

Most deployed neural networks this review is aware of are trained once and then frozen, because training it on something new tends to destroy what it already knew. This is catastrophic forgetting, and it is one of the principal obstacles to machines that keep working as the physical world around them changes, where the floor changes, the tool changes, and the failure is new. This report argues that the remedies with …

Read online →↓ PDF · 0.2 MBSurvey / Position · Preprint v0.1

The Energy-First Turn

Post-transformer, learn-while-inferring AI, the substrates built to run it, and the economics of changing what a unit of intelligence costs.

Computing's next constraint is not transistors, capital or talent: it is electricity. Data-centre demand is doubling this decade while physics permits computation millions of times cheaper than we practise it, because the bill is paid to move data rather than to compute it. This review maps the escape along two coupled axes (the substrates that execute relaxation natively, and the model classes that learn while infer…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v0.3

Verified Fluids: A Differentiable Computational Fluid Dynamics Stack for Physical AI, in the Browser

One open Rust stack spanning every solver paradigm, each checked against an analytic or reference oracle, differentiable end to end, and deployed self-verifying to WebAssembly.

Deployed robots do not run computational fluid dynamics in the control loop, and the reviews and vendor release notes this review located in 2026 place verification, rather than throughput or model size, as the layer they report as thin (sources listed in §1, Fact 3). We report a fluid-dynamics stack that treats verification as the primitive: every solver ships with an oracle it must reproduce before it is admitted. …

Read online →↓ PDF · 0.4 MBEngineering / Method · Preprint v1

One Digest: Bit-Identical Neural Computation Across Heterogeneous Compute Fabrics

Making a transformer forward pass compute the same bits in a browser tab, on Apple silicon, and on an NVIDIA GPU: the mechanisms that break it, the pins that hold it, and the measured price.

The same shader source, run on two GPUs, does not compute the same bits. Whether those 1-ulp differences compound into different behaviour over a full policy rollout is not measured here; this report measures the divergence where it originates, at the kernel and forward-pass digest, and pins it there. We report an engineering campaign that made a complete transformer kernel set (matmul, RMSNorm, LayerNorm, softmax, R…

Read online →↓ PDF · 0.4 MBEngineering / Method · Preprint v2

Touching Sound: Contact Audio and the Complementarity of Physical Perception

The sense that reaches through occlusion and into the sealed interior, and why sensor fusion is coverage, not redundancy.

Vision resolves an object's shape and a fingertip resolves its texture, but neither can tell a solid billet from a hollow shell, water from air behind a closed wall, or a seated bearing from a cracked one, and both go dark under occlusion or in the dark. Contact audio can. When a hand taps, scratches, or shakes an object, the vibrations it radiates carry precisely what light and touch cannot reach: material, internal…

Read online →↓ PDF · 0.3 MBSurvey / Review · Preprint v1

Impedance Matching as a Unifying Principle for Physical and Embodied AI: A Survey

One boundary condition, carried from a 1939 transmission-line calculator to the free-energy principle, and an account of where it stops being one.

A wave meets a boundary; whatever fails to match is reflected. This report surveys the extent to which that single statement organizes problems across the disciplines that physical and embodied AI draws on. It reviews five bodies of published work (the reflection coefficient and its conformal map onto the unit disk, Shannon capacity and the matched filter, the Bode–Fano broadband matching limit, Landauer's bound on e…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

Physical AI for Space Logistics and Transportation: Composing the Velocity Budget with One Estimation-and-Control Stack

A Δv-composition framework, an on-orbit non-cooperative capture, and a headless physics core the whole chain runs on.

Moving mass to, through, and around space is one continuous logistics chain: launch, orbital transfer, rendezvous, capture, servicing, assembly, and return. This report advances a research position: across that whole chain the decisive capability is autonomy carried on the vehicle, and the same estimation-and-control primitive recurs at every link, so a single autonomy stack composes across all of them. We make the a…

Read online →↓ PDF · 0.5 MBResearch / Method · Preprint v1

The Diffusion Layer: Accelerating the Spread of Silicon-Design Competency for Physical AI. A Survey and Research Position

Physical AI is bottlenecked less by fabs than by the people who can design for them. This report treats that competency as an innovation to diffuse, and the training as the technology that sets its speed.

Physical AI runs on custom silicon (for perception, for on-device inference, and for the energy budget of a machine that carries its own power). The size of that last term is measurable as joules per inference on a general-purpose part against joules per inference on a matched accelerator for the same model; this report does not compute it and treats the magnitude as out of scope, taking only the ordering from the so…

Read online →↓ PDF · 0.3 MBSurvey / Position · Preprint v1

The Human Layer: Competency at the Frontier of Autonomous Systems. A Survey and Research Position

As machines take the loop, the human role moves to the boundary. This report maps the competency that lives there, why automation erodes it, and how to build it deliberately.

Autonomy does not remove the human from a physical-AI system; it relocates the human to the system's boundary: to supervision, exception handling, calibration, and accountability. The competencies that boundary demands are different from the ones the machine displaced, and they are precisely the competencies that routine operation no longer exercises. This report surveys the long literature on that relocation, from F…

Read online →↓ PDF · 0.3 MBSurvey / Position · Preprint v1

The Bayou Air Corridor: Managed Low-Altitude Airspace over Houston's Greenways

Treating a continuous public greenway network as an AI-managed low-altitude volume, with flood search-and-rescue as the first mission and infrastructure inspection as the recurring one.

Houston's bayous are continuous public greenways: water and parkland that thread the entire metropolitan area. Measured by the standard ground-risk criterion for uncrewed flight, the population exposed beneath a flight path, the bayou corridors are among the lowest-risk ground in the city to fly over. Houston's mean population density is ~1,400 people/km² (2020 Census: 2,304,580 people over 1,651 km² of land). A 200 …

Read online →↓ PDF · 0.3 MBSurvey / Review · Preprint v1

Vision-Language-Interaction: Building a System to Interact, Not to Obey

A research position on making interaction, rather than task-completion, the objective of embodied agency, and metering every engagement with the world. Interaction is scored in joules per engagement; the same accounting, under disturbance, becomes the Institute's agency metric, J/VT (joules per viability-second held), defined and worked through for one engagement in Section 5. [In Section 5, add:] J/VT is the total energy charged to an engagement, E_compute + E_actuation, divided by the seconds the system holds its viability set under the disturbance. As an order of magnitude for a table-top manipulator: a policy step at [W_compute] over [s] and an actuation burst at [W_actuation] over [s] give [E] J for [t] viability-seconds, i.e. [E/t] J/VT. Every input here is named so a reader can replace it with the numbers of their own platform and recompute.

Contemporary embodied models are trained to map perception and language to action, and are optimized to complete named tasks. This report argues that task-completion under-specifies what a system with agency actually needs, because a policy is only ever prepared for the situations its objective named, and the world routinely presents situations no order anticipated. It develops a research position we call Vision-Lang…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

The Printed Body: How Far Additive Manufacturing Reaches into a Humanoid, Joint by Joint

Nine of the ten subsystems inside an electric actuator have a demonstrated additive route. The tenth is the drive die, and it must be fabbed.

A humanoid you can print is governed by one question: what can additive manufacturing actually do, joint by joint, and where does it hit a wall. This report treats an electric actuator as ten subsystems (structure, compliant transmission, flexure bearings, soft-magnetic iron, copper windings, permanent magnets, insulation, thermal path, sensing, and the power-electronics drive die) and asks, for each, whether a publi…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

The Surface That Pays Twice: Multi-Material Active Skins for Persistent Embodied Systems

A single persistence inequality, one physics-informed solver, and a skin that harvests and hides at the same time.

An embodied system that must persist (a stratospheric glider aloft for months, a Mars rover across a Martian year, a subsea drone that stays dark) is governed by a single inequality: the energy it harvests must meet the energy it spends, integrated over the mission. This report develops Multi-Material Active Skin Technology (MMAST), a research position in which the vehicle's outer surface is treated as its power plan…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

Spatial RF: Spoof-Resistant Sensing by Cross-Spectrum Corroboration

Forgeability sets trust; only keyed-coherent bands can veto; silence lowers the bar rather than blinding the system.

A single-band radio sensor is defeated by a single-band attack, which is the cheapest attack that exists. This report states a research position: a sensing fabric should span the whole usable spectrum and weight each band by how hard it is to forge, so that defeating the fabric requires forging a jointly consistent signature across every band at once. It reviews the primary sensing modalities the fabric composes (Wi-…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

Graph of the World: A Hybrid Property-Graph and Vector Substrate for Embodied Memory

One engine that traverses relationships and searches by similarity, on the device that has to act.

Perception fills a world model and physics makes it actionable, but the model has to live somewhere. This report argues that it lives as a graph. Every entity an embodied system knows (a part, a material, a surface, a constraint, another agent, a sensor reading) is a node, and every relationship between them is an edge. Because the physical world is continuous rather than discrete, each node also carries a vector emb…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

The Computable World Model: Retrieving the Physics an Embodied Agent Needs

Most of the physics an embodied agent needs is retrievable, not computable. A single cascade, over a single graph, from a milliwatt chip to a workstation. And the question it must serve ("will this action diverge before I commit?") is exactly what an energy certificate consumes: retrieved physics is the substrate under the certificate, not a world model for its own sake.

A world model is only actionable when the system can answer physical questions about it: will this part hold, will it fit, will it overheat, how far will it deflect. Full computer-aided design and physics simulation can answer such questions, but their kernels and solvers are too heavy to run inside a body's power budget. This report states and develops a research position, that of the CadFuture project of the Charlo…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

Spatial AI: A Living Gaussian-Splat Field That Predicts, at the Edge

A concrete definition: an explicit field of Gaussian splats, coupled to grounded dynamics, accounted in joules and run inside a robot's power budget rather than a datacenter's.

"Spatial AI" is the phrase the field reached for after "spatial computing," and it remains a buzzword in search of a definition. This report proposes a concrete one and defends it. A spatial model should hold the world as an explicit field of points, and Gaussian splatting is the representation that captures real geometry and appearance while remaining inspectable and editable. It should also be a world model that kn…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

The Contact Layer: Vision-Based Tactile Perception Where Line-of-Sight Ends

A camera under three colored lights becomes a geometry-and-force sensor, and the fingertip complement to a whole-body event skin.

Every remote sense goes to zero at the moment of contact. The last millimeter, the forces that hold a grasp, whether a surface is slipping, and the micron texture that distinguishes a bolt from a screw are not knowable from across a room. Touch is the perception layer that begins exactly where remote sensing ends. This report reviews the vision-based tactile fingertip, in which an elastomer gel deforms against an obj…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

OmniSense: The World Model as Projected Perception in a Volume

A projective account of the world model, in which where-and-what a system is emerges from a mesh of optical and radio-frequency nodes resolving a volume into occupied, empty, and unknown.

This report advances a projective account of a Physical AI system's world model. The model, including the system's own pose and identity within it, is treated not as an internal state read out from onboard sensors alone, but as a field defined over a volume by a mesh of sensing nodes: some carried by the system, some on peer systems, some anchored in the space. The mesh resolves the volume into three states. Positive…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

Provable by Construction: On-Device Lyapunov Certificates for Trustworthy Control

A controller is trustworthy not because it passed testing, but because its stability is proven, and the proof can ride with it.

Testing shows that a controller worked on the states it was tried on; it says nothing about the states it was not. A Lyapunov certificate says something stronger: that a closed-loop policy drives an entire region of the state space toward its equilibrium, by exhibiting a scalar energy that the dynamics can only spend. This report reviews how such a certificate is checked soundly and cheaply enough to run on the devic…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

MathGround: Pricing Every Decision in Joules, with a Replayability Class on Every Answer

A substrate that resolves each request at the lowest tier whose grammar covers it, meters the cost in measured energy, and stamps every claim with a type-enforced replayability class.

The cost of intelligence is energy, and most of what a system must decide is not a generation problem. This report describes MathGround, the substrate the remainder of the lab's work runs on. Two design commitments organize it. First, every decision is priced in joules and a request resolves at the lowest tier whose grammar covers it: a deterministic lookup, then a closed-form formula, then a sparse solver, and a sto…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

Energy-Native Computation and the Price of a Reproducible Bit: A Survey and Research Position

Order computation by the crossings between its energy and its state; find that the most efficient substrate is the noisiest; and price the reproducibility you must buy back.

In a low-power substrate, energy conversion and computation are frequently the same physical mechanism, described by two research communities that do not cite each other: what one calls harvesting, the other calls computing. This report organizes the field with a single axis (the number of domain crossings that survive between the energy source and the computational state variable), and surveys the resulting spectrum…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

A Thermodynamic Bound on Reproducibility in Energy-Native Computation, and the Crossings-Tax Spectrum

Collapse the crossings between source and compute state and efficiency rises but noise floods in. Reproducibility is bought back not with a conversion crossing but with a dissipation-built basin, at a price with a lower bound.

Energy-delivery-to-compute schemes are ordered by one axis: the number of domain crossings that survive between the energy source and the computational state variable, each crossing a conversion loss and a reproducibility checkpoint. Collapsing crossings raises efficiency and, along the same diagonal, forfeits the interfaces at which state could have been re-clocked or restored, so the maximally efficient substrate (…

Read online →↓ PDF · 0.3 MBCC0 1.0 · public-domain dedication

Microfluidics as a Shared Substrate for Physical AI: Co-Optimizing Heat, Charge, and Information in One Electrolyte Network. A Survey and Research Position

The seam where computation meets its energy budget: the same embedded fluid that must remove heat and deliver power can also carry a slow, local layer of computation.

An embodied system carries its power plant with it. At the edge, the dominant constraints on computation are not clock speed but the two currents that keep a chip alive: the heat it must reject and the power it must be fed. This report surveys a body of work in which those two currents already flow through a fluid (embedded microchannel coolant, and on-chip electrochemical flow cells) and observes that the same elect…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

Joule-Priced Co-Optimization of a Shared Microfluidic Substrate for Physical AI

One embedded electrolyte network, optimized jointly as coolant, power medium, and iontronic computer. Pricing any current alone is dominated whenever the couplings below are nonzero.

A single embedded electrolyte network is optimized jointly as coolant, localized power/regulation medium, and iontronic computational substrate. The three currents (heat, charge, information) share one flow and one dissipation budget. Under nonzero coupling (K1–K3, §4) the joint program below strictly dominates any pricing that optimizes cooling, power delivery, or computation in isolation. The disclosed object is th…

Read online →↓ PDF · 0.4 MBCC0 1.0 · public-domain dedication

Thermodynamic and Probabilistic Hardware for Generative and Embodied AI: A Survey

A generative model draws samples from a distribution. This report reviews hardware that draws them from physics instead of computing them first.

A generative model, reduced to its core operation, draws a sample from a probability distribution. A conventional accelerator does this in two stages: it computes the distribution with matrix multiplication, then draws from it with a random-number generator. Most of the energy is spent in the first stage, on a quantity that a random draw immediately collapses. Thermodynamic and probabilistic hardware removes the firs…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v1

Ternary and Low-Bit Neural Models, Their Training, and Their Silicon: A Survey

Restricting weights to {−1, 0, +1} removes the hardware multiplier. This report reviews what that buys, how such models are trained, and where the silicon stands.

Multiplication is the most expensive arithmetic operation in a neural accelerator, and the multiplier array is the circuit block that benefits most from an advanced fabrication node. Ternary weight quantization constrains each weight to the set {−1, 0, +1}. Under this constraint a weight–activation product reduces to a sign selection and the multiplier is no longer required. This report reviews the resulting body of …

Read online →↓ PDF · 0.5 MBSurvey / Review · Preprint v2

Post–von Neumann and Energy-Efficient Computing Paradigms for Physical AI at the Edge: A Survey

From the orbital gigawatt datacenter to the sub-microwatt microcontroller, with attention to how old each method is.

The dominant cost of contemporary artificial intelligence is energy, and most of it is spent moving data across the von Neumann boundary and performing floating-point multiplication. This report surveys the computing paradigms that reduce or remove those costs. It organizes about forty methods into nine families along two axes, paradigm and deployment scale, spanning energy-harvesting microcontrollers to orbital data…

Read online →↓ PDF · 0.4 MBSurvey / Review · Preprint v2