Where it breaks, not only whether
One score for each condition the robot works in: glare, smoke, dust, low light, parallax, occlusion, its own motion, its own lamp. You learn which conditions to trust it in and which to keep a person on.
Ragnarok tests and hardens the perception of autonomous systems so they read the world correctly and say when they can't. Inspection robots, camera AI, navigation without GPS and the people who supervise them, scored by condition, with the answer key sealed before any model runs.
One score for each condition the robot works in: glare, smoke, dust, low light, parallax, occlusion, its own motion, its own lamp. You learn which conditions to trust it in and which to keep a person on.
The readings it missed, the readings it invented, and the items it answered when it should have said it could not tell. Each one with the true value beside it.
The failures ranked by what they would cost on your site, and the test set kept so the next version has to pass the same items before it goes out.
The vendor picks the demo. The field does not.
A robot or camera AI is shown on the clean cases: a lit gauge, a dry lens, a still platform. Published accuracy rarely holds once the robot is walking, the lamp is on the glass and the air is full of dust. A demo does not test that. We build a sealed test for your site, with the answer locked before the robot sees an item.
The result is a number you can put in a procurement file, a safety case or a maintenance plan, and a test set the next software version has to pass.
Real base data with physics injected on top, or a rendered scene, so the true reading of every instrument and the true state of every scene is known before the robot sees it.
Answer keys are hashed and timestamped and the hash is published before any model runs. Nobody can tune to the test, including us.
Glare, smoke, dust, low light, parallax, occlusion, the robot's own motion and its own lamp. One score per condition.
Some items cannot be read. Planted zero-signal items count false alarms, and a system earns credit for declining only when declining was right.
Lasers and photonics, radar, acoustic and vibration, thermal and multispectral, navigation without GPS, and synthetic sensor test sets. What each sensor is good for, where it fails, and how we test it.
See sensors → Defense and public sectorSustainment, maintenance, navigation without GPS, force-protection detection, sensor trust and independent testing.
See defense →Your robot or model reads our synthetic instruments and scenes with exact truth. You get a scorecard by condition, the items it missed and the items it answered when it should have declined.
A demo is not a test. We run the vendor's system through conditions it has never seen and hand you the scorecard before you sign.
Your people supervising robot readings, with wrong calls seeded in. We measure what they catch, what they wave through, and how the hand-off holds as sensors degrade over a shift.
A public generator, sealed answer keys and baseline scores for several models, with scores by condition.
Faults planted in a synthetic plant so a robot's finds, misses and false alarms can be counted over a shift.
Scoring whether a robot knows it is lost, not only how far it drifted. A robot that drifts is a nuisance. One that drifts at full confidence is a hazard.
We test robots and the AI inside them. We do not build robots today. Every deployment we do ships with its own synthetic test environment and scorecard.
Gauges behind glare. Leaks in murky water. Dust underground with no GPS. Rebar and voids behind concrete. Machine health in the cold. Damage after a storm. The sensor changes. The question does not: did it read the world correctly, and did it say when it could not?
The robot will not tell you it was wrong. The incident report will.
A missed gauge is a shutdown or a release. A false alarm every shift trains the crew to ignore the real one. A robot that is lost and does not know it walks into something. A sealed test before fielding is a fixed fee and a few weeks.
Not today. We test them, and we build the test environments and scorecards that every deployment ships with. Hardware integration may come later.
No. Ragnarok tests and hardens the perception of autonomous systems so they read the world correctly and say when they can't. We do not build weapons or targeting systems.
Legged, tracked, wheeled and flying platforms, fixed cameras, and the models inside them. Cameras, thermal, lidar, radar, acoustic and vibration sensors, alone or fused. If it reads the world and someone acts on the reading, we can score it.
You see the conditions and the scoring rules in advance. You do not see the sealed items or the keys, and neither does anyone who funds the work. That is what makes the score worth having.
Send us one gauge photo your robot read, and the value it reported.
No charge for the first. Tell us the robot or model, the instrument or scene it reads, and the conditions it works in. We reply within one working day with how we would score it and what a test would cover.
Photos: inspection robot at a substation by Cvoellmy, CC BY-SA 4.0, cropped, this crop under the same license. Perseverance test rover, NASA/JPL-Caltech, cropped. Curiosity rover, NASA/JPL-Caltech/MSSS, cropped. Bomb-disposal robot, U.S. Army photo by Sandra Wilson, cropped. Lidar robot by S. Winkvist, public domain, cropped and edited.