The Charlot Lab.
The Charlot Lab works on the foundations of trustworthy embodied AI: systems that are provable, physically grounded, and cheap enough to run on the device. Two threads: Interface Engineering and Swap-2C Constrained AI.
Led by Dr. Charlot.
Publications
Every paper from this lab is a living one. Read it, run its model inline, ask it questions, call its method as a tool from Claude, navigate its place in the corpus, or let the Institute host speak it — a presence assembled from the paper's own concepts.
◆ The living corpusThe graphIn space · XR◆ Run as tools · MCP◆ Ferromotion · drive itThe Rust library
MathGround: joules, not tokens.
The cost of intelligence is energy, and most of what a system must decide is not a generation problem. MathGround is the substrate the rest of the lab's work runs on: every decision is priced in joules and carries a receipt, 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 stochastic model only as a last resort. Every claim carries a replayability class — deterministic, retrieved-and-cited, composed, or model-generated — enforced at the type level, so a generated answer can never be relabeled as ground truth. On commodity silicon the model tier costs orders of magnitude more energy than a cited composition; declining to invoke it unless the request demands it is where the joules are saved. The same substrate carries the lab's robot-policy harness: every frontier architecture behind one action space and one energy meter.
mathground.ai ↗Run a request: its latency is measured, its energy priced against real device power measured on silicon (macmon), and the receipt signed on-device with a real ECDSA key. Tamper with the replayability class and the signature breaks — provenance you can verify, not just assert.
In the field · the frontier is consolidating into world-action models that reason, generate, and act in one system (NVIDIA Cosmos 3, 2026), and into routing and cascades that fall back to a cheaper model when it suffices. MathGround takes the axis they leave open — energy as the unit, and a replayability class on every answer.
↓ Whitepaper · PDFRead online◆ Living paperTechnical Report TR-2026-06 · Institute for Physical AI @ BMI
Provable by construction.
A controller isn't trustworthy because it passed testing — it's trustworthy because its stability is proven. This is the other half of Swap-2C, the companion to the energy receipt: where the receipt prices a decision, this certifies a controller. A closed-loop policy is checked against a Lyapunov function by interval branch-and-bound — a real proof over a whole region of the state space, not a grid of samples — which returns either a certified region of attraction or a concrete counterexample state. Near the equilibrium the certificate is analytic, where the linear term provably dominates; everywhere else it is proven box by box, with the enclosures inflated so the result is sound rather than merely sampled. The whole check runs in the browser, on the device, in milliseconds — so a controller can carry its own proof to where it runs.
Tune the gains: the verifier proves V̇ < 0 by interval branch-and-bound and draws the certified region of attraction in gold — or finds the states where the pendulum's nonlinearity wins (red). Drop below kp ≈ g/L and the certificate correctly vanishes: the loop is unstable. Click the phase plane to release a trajectory.
In the field · learned Lyapunov certificates and certified training with branch-and-bound are the frontier of provable control (CT-BaB, FOSSIL). Swap-2C's line is its own: the certificate is cheap enough to run on the device, beside the joules receipt — a controller that ships both its cost and its proof.
↓ Whitepaper · PDFRead online◆ Living paperTechnical Report TR-2026-07 · Institute for Physical AI @ BMI
OmniSense: the world model as projected perception in a volume.
A Physical AI system's world model, including where and what it is, is a function of sensing projected over a volume by a mesh of nodes: some attached to the system itself, some on peer systems, some anchored in the space. The mesh resolves the volume into three states: positive space (occupied, reconstructed by line-of-sight optical: vision, stereo, LiDAR, event), negative space (confirmed empty, carved along each clear ray), and the unknown behind barriers, which spatial RF (Wi-Fi, UWB, mmWave) sees through to resolve, occupied and empty alike. A system perceives itself not from within but from the whole mesh at once, driving the unknown toward zero.
omnisense on GitHub ↗Sense → perceive → model → simulate → act, as a function of volume, location, and projection. Toggle Single agent ↔ OmniSense to see occlusion close.
↓ Whitepaper · PDFRead online◆ Living paperTechnical Report TR-2026-08 · Institute for Physical AI @ BMI
The contact layer: perception where line-of-sight ends.
Every remote sense goes to zero at the moment of contact. The last millimeter, the forces holding a grasp, whether a surface is slipping, the micron texture that tells a bolt from a screw — none of it is knowable from across the room. Touch is the perception layer that begins exactly where OmniSense ends. A vision-based tactile fingertip makes the trade concrete: an elastomer gel deforms against an object, a camera under three colored lights sees only color, and geometry is computed — not measured — on the device. Photometric stereo inverts the colored image into a field of surface normals; those normals integrate into a depth map; a grid of printed markers flowing with the gel reads shear and the onset of slip. An image sensor becomes a geometry-and-force sensor. This is the fingertip complement to the whole-body magnetic skin of the printed-body corpus: micron detail where dexterity needs it, an energy-accounted event skin everywhere else.
Press a shape into the gel and watch the pipeline: the raw RGB sensor image, the normals recovered by photometric stereo (n = M-1·[R,G,B]), the depth integrated from those normals (∇²z = ∇·g), and the marker field, where a gold stuck-core shrinks from the edge inward as the grip crosses the friction cone into slip. Drag the raw pane to apply lateral force — and watch the haptic-out trace, the vibration an actuator would replay as the finger slides across texture. From the reconstructed geometry alone it names the surface — a hex fastener, a knurled grip, raised dots — and with Auto-grip on, the fingertip applies the minimum force to hold, tightening the instant it senses incipient slip. Everything runs on the device.
Why it matters: a hand that doesn't drop things.
Slip is not an abstraction — it is the difference between holding an object and dropping it. Each contact can resist a tangential load only up to μ times its normal force; exceed that friction cone and the finger slides. Load the object below and watch a grasp fail, then let the reflex close it with the minimum pinch force.
Two fingertips, one object. Grasp holds while the friction capacity μ·2Fn covers the load; push the load past it and a contact's cone is exceeded — it slips red and the object falls. Auto-grip is the same slip-recovery reflex, now closing a grasp: it raises pinch force the instant the margin thins and relaxes it when the load eases.
In the field · vision-based tactile is the mainline of dexterous touch — GelSight's retrographic sensing and Meta's DIGIT 360 (with GelSight Inc.), an open hemispherical fingertip with ~8.3M taxels and multi-modal pressure, vibration, and shear. Two directions define the 2025–26 frontier. Touch is becoming a foundation-model problem — transformers trained by masked tactile prediction that transfer zero-shot across sensors (T3, AnyTouch). And event-based tactile now flags incipient slip hundreds of milliseconds to ~2 s before it turns gross — the same stick-to-slip transition this fingertip watches — with neuromorphic spiking networks near 94% on the no / incipient / gross split. Our recognition here is the transparent, on-device counterpart to those learned representations: features read straight off the reconstruction, no black box. And our line is the pairing of this micron fingertip with the printed-body event skin — full touch as an energy-accounted system, sensing and haptic return alike.
The learned version, live.
The recognition above reads hand-picked features off the reconstruction. The field has moved to learned representations — so here is that, at a scale you can watch: a small network trained by real backprop, in your browser, on raw depth patches with no hand-coded rules. The embedding clusters by texture as it learns; then it names a new press. It is the legible, on-device cousin of a tactile foundation model — same idea, 2,789 parameters instead of a billion.
A 100→24→12→5 network with real gradient descent. Watch the loss fall and the 12-D embedding separate into texture clusters (PCA-projected), then press Test to recognize a held-out touch. Reset the weights to watch it learn from scratch. Supervised on labeled example presses — honest few-shot, not the self-supervised, cross-sensor transfer of T3/AnyTouch — but the same principle: features are learned, not authored.
↓ Whitepaper · PDFRead online◆ Living paperTechnical Report TR-2026-09 · Institute for Physical AI @ BMI
Ternary Physical AI: a policy with no multiplies, on silicon anyone can make.
The cost of a decision is energy, and most of that energy is spent multiplying. A multiplier is the densest, most node-hungry block on an AI chip — the reason a good policy seems to demand a leading-edge fab. Ternary weights {−1, 0, +1} make the multiplier disappear: every multiply becomes an add, a subtract, or a skip. BitVLA showed a 1-bit Vision-Language-Action model can match a full-precision baseline at 11× less memory and 4.4× lower latency — the policy did not get worse, it got cheaper. Once inference is adds and lookups, the hardest part of the chip is gone, and a mature, high-yield, sovereign node becomes enough: 65 nm ternary compute-in-memory already reaches ~20 TOPS/W. This is the Swap-2C thesis in silicon — intelligence priced in joules, small enough for the body it lives in, on silicon that is not scarce.
A small policy — observation patch → action — trained live, then quantized. Toggle FP32 → Ternary {−1,0,+1} → Log and watch the ledger: the weights go 20× smaller, the multiplies go to zero, the energy falls, and the accuracy holds. It stays accurate because it is trained ternary (quantization-aware, straight-through), not squeezed after the fact. The silicon panel shows the payoff — no multiplier array means a mature node will do.
In the field · the energy problem is concrete — in the datacenter it is a grid-scale line item, on a robot it is the whole battery, and most of those joules are burned in the multiplier. Three camps chase it. Ternary lives in the models, and the models keep moving: Microsoft's BitNet b1.58 (with bitnet.cpp), its 2026 Sparse-BitNet, and BitVLA for embodied policies. The silicon lags: the most genuinely multiply-free ternary parts — BitROM (65 nm), TENET (28 nm), VitaLLM (TSMC 16 nm, 66 mW) — are research, mostly post-layout simulation rather than taped-out chips, and often heterogeneous, a conventional integer core still doing attention. Log-math is earlier still: multiply-becomes-addition (BitEnergy's L-Mul) is an algorithm ahead of its hardware, and the clearest log-math chip startup, Lemurian, has left silicon for software. Meanwhile the mainstream routes around both — into low-bit floating point (OCP's MXFP4, NVIDIA's NVFP4, Huawei's HiFloat8) — conceding that the number format is the lever while declining to pull it all the way. The models went 1-bit before anyone shipped 1-bit-native hardware; no fully-ternary or log-math ecosystem ships yet. That gap — multiply-free intelligence on mature, sovereign silicon, priced in joules — is what this topic works toward.
↓ Download whitepaper · PDFRead online◆ Living paperSurvey · preprint v1 · open scholarly use
Thermodynamic Physical AI: when sampling is the compute.
Half of a policy is deterministic — a feedforward pass, which ternary makes multiply-free. The other half is stochastic: sampling an action, a plan, a belief about a half-seen world, the denoising steps of a diffusion policy. And a generative step, underneath, is just drawing from a probability distribution — which a digital chip does the hard way, computing the distribution with matrix-multiply and then sampling it. Thermodynamic hardware inverts the order: it builds the distribution into a physical system and lets thermal noise fall into it, sampling directly. The noise a digital chip spends energy fighting becomes the computation. Extropic's sampling units and Normal Computing's thermodynamic chip claim orders of magnitude less energy for exactly the generative workloads Physical AI leans on, and networks of p-bits already do the same a million at a time. Pair it with ternary and the edge stack sits at its energy floor on both halves — deterministic and stochastic, each done the cheapest way physics allows.
Walkers run Langevin dynamics on an energy landscape — a stand-in for a thermodynamic sampler drawing from a multimodal belief. Turn the temperature down and it freezes into one mode (the mixing–expressivity tradeoff); turn it up and the structure washes out; in between it samples the true distribution. The energy panel is the point: sampling by physics runs orders of magnitude below matmul-then-sample, and the p-bit strip is the free randomness that does it.
In the field · thermodynamic computing is young but real. Extropic builds sampling units that draw from energy-based models in silicon (XTR0 prototype, thrml library, Z1 for 2026), claiming ~10⁴× less energy than a GPU on generative workloads; Normal Computing taped out CN101, a first thermodynamic chip. Underneath sit p-bits — stochastic magnetic tunnel junctions fluctuating at GHz — now networked a million at a time doing >10¹² Gibbs flips per second, and denoising thermodynamic models that borrow diffusion's trick to beat the mixing–expressivity wall. The claims are large and the silicon is early — the honest read is that the deterministic path (ternary) and the stochastic path (thermodynamic) are the two lanes where the energy of intelligence actually falls, and the edge is where they have to meet.
↓ Download whitepaper · PDFRead online◆ Living paperSurvey · preprint v1 · open scholarly use
The open field of computing: every way to compute, and how old it is.
Von Neumann's architecture — fetch, compute, store, repeat — is one idea among dozens, and not the most efficient. Beneath it is a whole field: numbers that make the multiplier vanish, memory that computes where it sits, physics that samples for free, light that multiplies at no cost, matter that solves a problem by settling into its lowest energy. This atlas surveys them, from the orbital gigawatt datacenter down to the sub-microwatt microcontroller. The striking thing is the dates. Balanced ternary shipped in the Setun computer in 1958. The memristor was predicted in 1971. The Ising model is 1925; cellular automata, the 1940s; reversible computing, 1961. These are not moats — they are deep, published prior art, an open commons anyone can build on. Which reframes the question: the future of computing is not the discovery of one new method — every method has already been thought of and tried. It is the convergence of the open ones into a single machine that spends energy only where physics demands it.
Every node is a real method with a real history; the year is when its prior art begins. Read the map left-to-right by scale — a batteryless microcontroller running on ambient energy, out to a gigawatt solar datacenter in orbit — and top-to-bottom by paradigm. Filter by family; click any method for its principle, its energy or math advantage, and the century-deep lineage it rests on. We go deep on two of these — ternary for the deterministic path and thermodynamic for the stochastic — because their convergence is the edge stack.
In the field · the map's edges are moving fastest. At the largest scale, orbital datacenters left slideware behind — Starcloud trained a model in orbit in December 2025, and Google's Project Suncatcher aims solar-powered TPU satellites at 2027. At the smallest, energy-harvesting microcontrollers compute with no battery at all. In between, every unconventional paradigm — in-memory and memristive, neuromorphic, photonic, thermodynamic — is closing on practicality at once. The Institute's position is the plain one the dates imply: none of this is anyone's private future. It is an open commons, and the work is to converge it well.
↓ Download whitepaper · PDFRead online◆ Living paperSurvey · preprint v1 · open scholarly use
Spatial AI: the world as a living field of points.
"Spatial AI" is the term the field reached for after spatial computing, and it is still a buzzword in search of a definition. The lab proposes a concrete one. A spatial model should hold the world as an explicit field of points — Gaussian splats, the representation that captures real geometry and appearance — and it should also be a world model that knows how that field moves: the true dynamics of the scene, not a frozen scan. Spatial AI is the mash-up of the two, a living splat field that predicts. Most of the field builds this in the datacenter; the Swap-2C constraint is the opposite — it has to run at the edge, on the low-power device that lives in the world it is modeling. The representation stays explicit and inspectable, the dynamics stay grounded, and the whole thing is accounted in joules on the MathGround substrate. A world model the size of a robot, not a server farm.
A live field of Gaussian splats — depth-sorted and rendered in real 3D on the device, explicit and inspectable, and predicting its own motion: the orange ghost is the field's forecast +Δt ahead. Drag to orbit. Synthetic scene, illustrative.
In the field · World Labs frames a world model as renderer → simulator → planner; the running Gaussian-splat world models (GaussianWorld, GEM) already predict an evolving volume, and on-device splatting is now real (Mobile-GS, ~1000 FPS at a few MB). The open axis is dynamics at the edge — a splat field that predicts, on the low-power body. That is what Spatial AI targets.
↓ Whitepaper · PDFRead online◆ Living paperTechnical Report TR-2026-10 · Institute for Physical AI @ BMI
The Computable World Model.
A world model is only actionable when the system can model the physics of it: will this hold, fit, overheat, deflect. Real CAD and simulation are too heavy to run where the body is. CadFuture's claim is that the physics an embodied agent needs is mostly retrievable, not computable. Every query walks one cascade, lookup then closed-form formula then sparse solver then a model only as a last resort, over a single graph that carries geometry, the physics fields, tolerances, live-twin state, and the agent's own interaction. The same full engineering model runs from a sub-five-milliwatt neuromorphic chip to a workstation; execution adapts, the engineering truth does not.
cad-future on GitHub ↗Illustrative. Physics queries stream in; most resolve at the LUT shelf for about a picojoule, and only a trickle fall through to a solver or model. The research effort is CadFuture, a Charlot Lab project.
↓ Whitepaper · PDFRead online◆ Living paperTechnical Report TR-2026-11 · Institute for Physical AI @ BMI
Graph of the World.
Perception fills a world model and physics makes it actionable, but the model has to live somewhere. 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; every relationship is an edge. And because the world is continuous, each node carries a vector embedding, so a system can both traverse relationships and search by similarity in one engine, on the device. Graph of the World is that substrate, the hybrid graph and vector store that holds what a system perceives and reasons over. The research effort is hyperdb.
hyperdb on GitHub ↗A property graph and vector search (HNSW) in one engine; queries traverse relationships and find nearest neighbors at once.
↓ Whitepaper · PDFRead online◆ Living paper◆ Explore the live corpus graphTechnical Report TR-2026-12 · Institute for Physical AI @ BMI
Spatial RF.
The cheapest way to spoof a sensor is to attack one band, so sense across all of them. Spatial RF fuses the whole spectrum — mmWave vitals and fall detection, C-band through walls, authenticated UWB ranging, Wi-Fi CSI, BLE, mesh tomography, sub-GHz perimeter — weighted by how hard each band is to forge. Low-trust ambient bands corroborate but can never veto a real event; only a high-trust, keyed-coherent band can. Silence is a fault, not a gap: a dead or jammed node lowers the bar rather than blinding the system. Defeating the fabric means forging a jointly-consistent signature across the entire spectrum at once, which is physically incompatible. The research effort is Sentinel.
sentinel on GitHub ↗A real event corroborates across bands; a spoofed band falls off the manifold; a dead band fails loud. The fabric cycles: monitoring → fall → spoof → fault.
↓ Whitepaper · PDFRead online◆ Living paperTechnical Report TR-2026-13 · Institute for Physical AI @ BMI
The surface that pays twice.
An embodied system that has to persist — fly for months, run for a Martian year, stay dark — is governed by a single inequality: harvest must meet demand, integrated over the mission. MMAST treats the vehicle's skin as the power plant and reads every gram against that ledger. The same multi-material stack that harvests also reshapes the signature: the radiative-cooling film that cools the cells is the IR-stealth layer; the metasurface that lifts photovoltaic gain is the radar-absorbing coat; the TENG skin that scavenges vibration is the distributed structural-health sensor. Every module pays back twice. It is one physics-informed solver over three orthogonal axes — vehicle archetype, medium, and surface module — so the same energy balance holds for a stratospheric glider, a subsea drone, and a Mars rover. The simulator is built on CadFuture, the lab's computable-world-model engine, runs headless for parameter sweeps, and renders in the browser through WebGPU.
physics-mmast-sim on GitHub ↗Each archetype balances its harvest stack against its demand stack; cross the break-even line and the vehicle is persistent. The fleet cycles: HALE solar UAV → cloud-carrier airship → subsea drone → Mars rover.
↓ Whitepaper · PDFRead online◆ Living paperTechnical Report TR-2026-14 · Institute for Physical AI @ BMI
The printed body.
A humanoid you can print — the frame, the transmission, the actuators, the skin — is governed by one honest question: what can additive manufacturing actually do, joint by joint, and where does it hit a wall. The lab's answer is that nine of the ten subsystems inside an electric actuator already have a demonstrated additive route — structure, compliant transmission, flexure bearings, soft-magnetic iron, copper windings, bonded magnets, sensing — and the tenth, the drive die, is the one part that must be fabbed, not printed. So the robot is not a printed shell with bought muscles: every joint is one axial-flux actuator, AXF-1, re-sliced per joint from a single design — same poles, coils, magnets, and process, only the diameter changes. The cost of full printability is legible rather than hidden: bonded-magnet torque runs about half of sintered, so a lean 14-DoF walker carries roughly ten kilograms of actuators, and the sovereignty tax for printing the legs instead of buying them works out to about +38% of actuator mass in the design study. Every joint is priced in joules on the MathGround substrate, and every number traces to one cheap coupon — the remanence of a printed magnet. The research effort is the Free Humanoid Corpus, released CC0 as prior art.
Browse the Free Humanoid Corpus →on GitHub ↗The body prints from the feet up; each joint is an axial-flux actuator with a printed body and one fabbed drive die at its core. The corpus is the CC0 chain: capability matrix → actuator family → magnet gate.
In the field · open printable humanoids are real — Berkeley Humanoid Lite and ToddlerBot both print the body and buy the actuators, and MADE3D prints whole electric machines — for wind turbines. The open axis is the one nobody occupies: a printed actuator at humanoid load, integrated and accounted in joules per joint.
↓ Whitepaper · PDFRead online◆ Living paperTechnical Report TR-2026-15 · Institute for Physical AI @ BMI
VLI: built to interact, not to obey.
A digital system with agency cannot only take orders. To act in the world it has to interact with the world — and with everything in it: people, objects, surfaces, forces, signals, other systems. Vision-Language-Interaction makes interaction itself the objective, not task-completion and not obedience. An order-following system is only ever ready for the orders it was given; a system built to maximize interaction is, by construction, assembling the repertoire it needs to act in any situation it meets — interaction is how it meets the world and learns what it can do there. We have watched machines move with intention for a century, in the films and the anime and the toys on the shelf, long before one could be built; what is new is that interaction, not command, can now be the thing a system is built to maximize. The lab's effort — Ambit — is its own, and it is grounded: every interaction runs on the MathGround substrate, so each decision to engage the world carries a joules receipt and a replayability class.
ambit on GitHub ↗The agent reaches out to everything in its world — people, objects, surfaces, forces, signals. As interaction coverage grows, so does its readiness for situations it was never given. Order-following touches only a sliver of the world.
In the field · the interaction-first idea is surfacing in pieces — concurrent see-hear-speak-act (VITA-E), and policies that learn from their own experience rather than pure imitation (π*0.6). Ambit's line is its own: interaction as the objective itself, grounded and metered on MathGround.
↓ Whitepaper · PDFRead online◆ Living paperTechnical Report TR-2026-16 · Institute for Physical AI @ BMI
Space logistics and transportation.
AI, Physical AI, and Embodied AI are opening new efficiencies, new materials, and new optimizations across space transportation and the management of space logistics. Reaching a destination in space is a velocity budget, and the whole chain, from launch through orbital transfer, rendezvous, capture, servicing, assembly, and return, becomes tractable when reusable, shared infrastructure and autonomous in-space transport pay that budget down. A momentum-exchange tether catch, a launch loop held steady by its controller, an electromagnetic sled handoff, a rendezvous with a tumbling target: each is an estimator resolving state and a controller holding a proven corridor, run on the vehicle at the edge. This track is the research into what those capabilities are and how to test, apply, and deploy them, so the space economy can expand, open, and decentralize.
Browse the instrument suite →orbital-logistics on GitHub ↗Pick a destination and compose the chain: reusable infrastructure and autonomous in-space transport pay down the velocity budget the launch vehicle would carry alone. The mass-ratio math is the exact rocket equation; the Δv budgets are representative.
In the field · autonomous rendezvous and proximity operations already run in orbit. Astroscale's ADRAS-J approached a non-cooperative three-tonne rocket body, held station within fifteen meters, flew around it, and validated autonomous collision avoidance; Northrop Grumman's SpaceLogistics Mission Extension Vehicles docked with client satellites in geostationary orbit to extend their lives. These are the first links of an in-space logistics layer. The research here is the autonomy that lets one estimation-and-control stack serve every link in the chain, from a launch handoff to a non-cooperative capture, priced in Δv and joules and carried on the vehicle.
↓ Whitepaper · PDFRead online◆ Living paperTechnical Report TR-2026-17 · Institute for Physical AI @ BMI
Two threads.
What the lab works on.
Clean seams between parts
The engineering of the interfaces between the pieces of an embodied system (hardware, software, sensors, and subsystems) so modular systems compose without bespoke glue.
Provable, grounded, cheap to run
AI grounded in physics and math rather than language: closed-form primitives, controllers traced to a Lyapunov function, and a picojoule energy receipt on every call. Behavior is provable and the power cost is known before deployment. It runs in a WebGPU browser in under 400 KB.
Open Interface Engineering.
One thesis runs through the topics, the threads, and the tracks: the lab is Open Interface Engineering in practice. The hard failures in complex systems happen at the interfaces between parts, and the binding cost of AI is energy.
Where systems fail
The hard failures in complex systems happen at the seams between parts. The lab engineers those interfaces so systems compose and stay verifiable, instead of fusing into bespoke glue.
Physics, not vibes
Behavior is grounded in physics and math and traced to a certificate, so what a system will do is known before it runs, not discovered after.
Account in energy
Every computation has a thermodynamic cost. The lab measures and minimizes it in joules, the honest unit for AI that acts in the physical world.
Five tracks, each a live site.
The lab's research runs in the open. Each track has its own thesis, a fancy visual, and a working site you can use.
MathGround
The Swap-2C runtime: physics- and math-grounded primitives, closed-form control, and a picojoule energy receipt on every call.
Open the research track →Pattern-Lang
Recognizing patterns by naming, composing, and verifying them over a small finite lexicon, instead of learning them from billions of examples.
Open the research track →Play Dimension
A strategy game whose mechanics map directly to ideas from AI and neuroscience.
Open the research track →openIE CAD
CAD/ECAD that AI agents can actually drive: geometry, PCB, optics, simulation, and manufacturing through one MCP endpoint, in a browser-native WebGPU viewer.
Open the research track →OpenIE Compute
The Periodic Stack of Computation: every computational primitive mapped by consistent axes, including its thermodynamic cost in joules.
Open the research track →Build this with us.
The lab takes on students through internships, and works with investigators across the Institute.