Snap it together. Then make it think.
A magnetic or snap connector collapses three separate maker skills — structure, wiring, and firmware — into one self-aligning click. It removes the exact barriers that stop beginners: the soldering iron, the reversed polarity, the mounting bracket. The budget you save goes where it matters — to behavior. That is why connection systems are the on-ramp to physical AI, and why LEGO, magnetic blocks, and the four-pin sensor bus all belong in the same story.
From a click to a policy.
The same rig, made more intelligent at every rung. The first five are shipping-grade today — you can teach them now. The sixth is real but not yet solderless, not yet browser-native. That last step is the one we build.
Snap a cause to an effect
Cubelets · littleBitsteach it nowBlocks that pass a value across magnetic faces — sense, think, act, no code. You hold the loop of an embodied agent in your hands. Real intuition; a hard ceiling — no camera, no learned policy.
Snap on real senses
Qwiic ≡ STEMMA QT · Groveteach it nowThe USB of sensors: click an IMU, a time-of-flight depth imager, or a motor driver onto a board with no soldering and no polarity to get wrong. Real spatial perception on a four-pin bus.
Write the rule by hand
MakeCode · Blocklyteach it nowIf it tilts past here, turn. The explicit policy — written out, so that when you replace it with a learned one you know exactly what changed.
Train a model on the device
micro:bit CreateAI · Grove Vision AIteach it nowRecord your own motion or your own objects, train in the browser, run the model on the board. The smallest real physical AI you can snap together today — and CreateAI already streams over Web Bluetooth.
Let a vision policy drive
HuskyLens → micro:Maqueenteach it nowA no-code recognition camera steers a rover. Real as a product, shallow as learning: you consume the model, you do not own its training.
Teleoperate it, then let it teach itself
LeRobot SO-101 · and our stackthe gap we buildDrive the rig, record the episodes, train a policy, redeploy it to the same hardware. The real thing — and today it is neither solderless nor browser-native. Closing that, in the browser, is the gap we build into.
Climb this ladder for credit.
Every rung on this page is a graded, in-browser lab: write the sense→think→act loop, decode a sensor off the bus, watch a learned block out-drive your hand rule, switch a magnet that holds for free, and clone an expert from its own demonstrations. Finish the five and claim a verifiable certificate.
What actually carries what.
An honest map of the interfaces. Three of these carry power and data over a self-aligning magnet; two are the solderless sensor bus; one is a wireless robot protocol we already drive from the browser. Each names the one sensor it cannot carry — because that gap is where the design lives.
Qwiic ≡ STEMMA QT
I2C · JST-SH 1.0 mm
CARRIES IMU, ToF depth, motor & servo drivers
Physically and electrically identical — cross-plug them freely. A camera is the one sensor this four-pin bus can't carry.
In the Atlas →Grove
I2C / UART / analog · JST-PH 2.0 mm
CARRIES the widest catalog, including vision-AI blocks
Not pin-compatible with Qwiic, but a hub bridges them. The most versatile of the solderless buses.
In the Atlas →LEGO Powered Up (LWP3)
Bluetooth LE
CARRIES encodered motors, colour / distance / force
Openly documented and drivable from a browser — we already have the driver. No camera and no on-hub ML: LEGO is the body; the mind runs in the browser or on a Raspberry Pi.
In the Atlas →Magnetic pogo-pin
power + data, self-aligning
CARRIES anything — it is the interface, not the payload
Spring pins for contact, a magnet for retention: 100,000+ mating cycles, polarity-safe, no aim required. MagSafe is the one you already own; a drone charging dock is the autonomy version.
In the Atlas →Cubelets magnetic face
three conductors · ground ring, power ring, data pin
CARRIES one value, 0–255, one direction
The clearest physical picture there is of a keyed, polarity-safe port — and of the sense → think → act dataflow itself.
In the Atlas →Electropermanent magnet
flip a permanent magnet with a pulse
CARRIES a bond you switch on and off with zero standing power
The substrate of reconfiguration and self-assembly — M-Blocks and SMORES-EP in the lab, FluxGrip on the shelf. A body that changes its own attachments makes morphology part of the policy.
In the Atlas →Four questions, honestly scored.
A kit reaches physical AI only if it can carry a rich sensor, stream that off-board to a policy, let its morphology be rebuilt, and expose the sense → think → act loop. No off-the-shelf kit scores full marks — the magnetic power-and-data kits and the camera-and-policy kits are still two disjoint sets. That disjointness is the opening.
| Kit | Camera / rich sensor | Streams to a policy | Reconfigurable body | Shows sense→think→act |
|---|---|---|---|---|
| Cubelets classroom-real | ◐ | ○ | ● | ● |
| littleBits legacy supply | ○ | ○ | ● | ● |
| Circuit Cubes classroom-real | ○ | ○ | ● | ◐ |
| micro:bit + CreateAI classroom-real | ○ | ● | ○ | ● |
| Grove Vision AI · reCamera classroom-real | ● | ● | ○ | ◐ |
| HuskyLens + micro:Maqueen classroom-real | ● | ◐ | ○ | ● |
| Robo Wunderkind · Tinkamo classroom-real | ● | ● | ● | ● |
| LEGO + Build HAT + Pi Cam classroom-real | ● | ● | ○ | ● |
| LeRobot SO-101 frontier | ● | ● | ○ | ● |
● yes ◐ partial ○ no
The deploy channel is a tab.
The browser talks to real hardware over Web Serial and Web Bluetooth — no toolchain to install. Our sim-to-real layer already carries a learned policy out of a page and into a real board. The connection system is how it lands in the world.
A real LEGO hub
Our driver speaks the LEGO Wireless Protocol to a SPIKE, Technic, or Robot Inventor hub — pick it in the browser's Bluetooth prompt and stream motor commands straight to it. LEGO is the fastest first real-hardware rung there is.
A CyberBrick, an ESP32, a micro:bit
The Maker Studio's programming bench flashes and drives microcontrollers over Web Serial and BLE-UART from the tab you are reading this in. Seventeen hardware targets, one deploy channel.
Perceive in the browser, act on the metal
Bluetooth adds milliseconds and jitter, so the fast loop stays on the hub or a co-located Pi while the browser runs perception and streams setpoints. Knowing why is itself a physical-AI systems lesson, not a limitation to hide.
Sense, think, act — and imagine.
The whole arc of physical AI, each stage a model you train yourself — on-device, in a browser tab, at zero cost. Four live labs: snap-together perception, a policy that out-learns your hand rule, an arm you teleoperate into a skill, and a world model you can dream a policy inside.
Teach it to see
Show a camera a few examples, train a tiny classifier in your browser on-device, and watch it label the live feed — exactly what a $16 vision module does. No cloud; nothing leaves the page.
Open the interactive →
02 · THINKSense → Think → Act
A block graph drives a rover: tune the think block by hand until it reaches the goal, then flip on a hidden hardware quirk and watch the same rule spiral away. Swap that one block for a learned one, train it on-device, and it finds the fix from reward alone.
Open the interactive →
03 · ACTThe SO-101 loop
Teleoperate a pick-and-place arm, record demonstrations, clone them into a policy with in-browser behaviour cloning, then watch the policy do it alone. The $100 SO-101 and LeRobot loop, at zero cost. Stream the same policy to a real arm over Web Serial.
Open the interactive →
04 · IMAGINEImagine, then act
An agent plays at random, learns a world model of the hidden dynamics, then trains a policy entirely inside that model — in imagination, zero real steps — and acts for real. The 2026 world-action-model idea, on-device.
Open the interactive →
The credentialed way through all four: the Perception to Policy course →
The magnet, switched
The connector that makes a body reconfigurable: an electropermanent-magnet latch you pulse on, hold with zero standing power, and pulse off. Drag a module into the capture basin and it self-aligns. Grounded in the real force-and-energy numbers.
Where these systems sit next to the boards that run them: Digital Brains → · the whole landscape: the Maker Atlas →
Build a policy — then judge it.
The four verbs build a policy. But how do you tell a good one from a bad one, and improve it, without hand-writing the reward? You learn the reward. Compare a few behaviours, say which you prefer, and a model learns to score any behaviour the way you do — the RLHF-for-robots idea (Robometer, TOPReward, LeRobot's 2026 reward-models API). A naive rule that only asks did it reach ties every success; the learned reward tells a smooth reach from a jerky one. But a reward only ranks what already happened — it never certifies that the plan you are about to commit won't diverge. Ranking a behaviour and guaranteeing one are different objects: a reward ranks, a certificate verifies (energy-based control even shows failure recovery a certificate would explain rather than merely score — EBT-Policy, 2025).
Learn the reward
Compare pairs of behaviours, click the better one, and train a reward model on your preferences — then watch it rank every behaviour, on-device. The same idea that ranks real robot policies in 2026.
What the loop actually runs on.
Every system above assumes a board, a bus and a power budget. This is that layer: what the compute costs, what the sensors draw, and what a real build has to fit inside.
A policy isn't free — it spends energy.
A controller can track perfectly and still be the wrong answer: every correction costs effort, and effort costs joules. This is the axis the Institute optimizes — performance per watt, not benchmark points. The figure below turns that tradeoff into a dial you can feel.
The cost of control
Turn a controller's gain up: faster tracking, more effort. Set the energy price and the cheapest gain moves. Every value is hover-inspectable — a static plot turned into a live instrument, the template every figure here is moving to.