Before a delivery robot takes a job, we crash-test it 120 times in a physics copy of that exact route. How often it fails becomes the price. If it really fails, we replay the run to find the cause, and PayPal pays the claim in seconds.
Open the live app Watch the film Code on GitHub8:47 pm, the first snow of the year. Maya’s little bakery has one delivery left: a birthday cake for a seven-year-old. Her robot, Bolt, slides on the icy driveway into the road. The cake never arrives.
Robots now work on real sidewalks, steps and factory floors, and they fail. But nobody can insure a single robot job, because nobody can say how risky this robot, on this route, tonight really is. So the cover doesn’t exist, and small businesses like Maya’s carry the whole loss.
A MuJoCo physics copy of the real route: the slope, the curb, the weather odds. Describe it or upload a photo; Gemini builds it.
Every run is a different day: dry, rain, ice, a stranger bumping in. The failure rate is the risk.
Gemini plays the villain: suggests how Bolt fails; we search, shrink and re-check it on 32 fresh runs.
Bolt practises on its failures and ships only after 3 strict checks. A safer robot pays less.
Replay the exact run, removing one thing at a time. Covered cause → PayPal payout. No paperwork.
of 120 simulated days, Bolt fails the icy back-door route
to insure a $400 delivery: risk × cover × 1.25 + $0.30
after Bolt retrains and passes the locked 16-test gate
confirmed failures found by the AI adversary vs blind random attacks, same budget
# premium: the customer pays POST /v2/checkout/orders intent=CAPTURE PayPal-Request-Id = policy hash # idempotent reference_id = policy HMAC signature POST /v2/checkout/orders/{id}/capture # claim: the robot owner is paid POST /v1/payments/payouts sender_item_id = claim id # every event: verify-webhook-signature
The price is a PayPal Order the customer approves. The policy’s cryptographic signature is stamped on the order, so every payment traces back to the exact risk we measured. A covered failure triggers a PayPal Payout to the owner. Tested end to end on the PayPal sandbox, including a real $400 sandbox payout.
A brand-new category of payments: robot jobs. Illustration: 1 million robot jobs a month × a $1.50 average premium = $1.5M a month of new volume through PayPal Orders, plus every claim paid out with PayPal Payouts.
The real app on the PayPal sandbox. The robot’s physics was run ahead of time, so every answer is instant. Pricing, the AI adversary log, the gate, claims and every PayPal call happen live.
Keep “Maple St back door”, press Run 120 days and price it: 120 simulated days become a price.
Start the attack, then Practise and check. Watch the gate, re-price: it drops.
Pay with PayPal opens sandbox checkout. Log in with the sandbox buyer from the Devpost testing notes.
Set weather to ice, Do the delivery, then Make a claim: replay decides, PayPal Payout pays.
Free hosting sleeps when idle: the first visit can take about a minute to wake up.
The premium: create, buyer approval, capture. Idempotent request id, policy HMAC as reference id.
The claim: paid to the robot owner in seconds, claim id as sender item id. Webhooks signature-verified.
The adversary’s hypotheses, a photo or description turned into a route, and the claim adjuster’s note.
The physics twin: a six-wheel robot, curbs, ramps, tilted aprons, ice, bumps, cargo, motor sag.
The live ledgers: policies, claims, every attack and every robot version, side by side.
Hosts the live demo from the repo’s render.yaml, one click to deploy your own copy.
Also: FastAPI, three.js (3D replays and the film), NumPy.
A 3-minute film rendered from the real physics recordings, then the live app running on the PayPal sandbox. Watch it on YouTube