The release flywheel of physical AI

Every deployment makes your robots better.

Your team says what the robot should do and which limits it must respect. We're building the flywheel that does the rest, starting with a shadow engineer that takes you from field failure to validated redeploy.

What goes in
Logsevents, errors
ROS bagslidar, camera, IMU
Telemetryflights, fleet runs
Configparameters, sensors
Robot codecommits, releases
Site videoscustomer sites
Site scanspoint clouds, maps
Jira, Slacktickets, conversations

Your shadow
engineer

Every scenario, every releaseHeld at the gate · 213/214
PassNew from the fieldStopped by the gateNo scenario yet

Every failure in the field becomes a test that runs on every release.

So your team focuses on what's new, and what you fixed stays fixed.

Acme Robotics/quad_x/InboxExample data
quad_xDrone · ROS 2
Your shadow engineer

Post-processed yesterday's site test

3 relevant patterns detected across 14 flights.

What it took into account
  • Flight logs14 MCAP files
  • TelemetryPosition, IMU, setpoints
  • Robot codeRelease 4.3, 12 commits
Open report
  • Nightly suite passed212 of 212
Yesterday's site testNorth site · 14 flights
Unstable handover3 flightsLidar dropout2Person too close1
What happened
The drone went unstable when the pilot handed back to autopilot. It jumped 1.1 m.
What it didDerived a new scenario to cover this transition.
Open scenario
Teleop to autopilot handoverNew
Scene
North siteReused from the last site scan
What happens
Pilot takes over, then hands back24 variations
Pass when
No jump above 0.8 0.5 mNo acceleration spikesLimits tuned on the 11 clean flights
Real flight against simulationRound 1
41%match with the real flight
  1. Gazebo41%
  2. MuJoCo68%
  3. IMU model fitted to the flight84%
  4. Noise and a sensor fault injected96%
The simulation now fails the way the drone failed.
Incident
NL-2481Unstable handover, reproduced in simulation.
Copy ID
Paste it into your coding agent. Through the Nonlimi MCP it reads everything found so far, and acts on your code.
Release suite214 scenarios
Running4.3.1
The whole suite checks the fix.
It runs on every change from now on. Your next site test can go to what's new.
Sign release
Where this is going

Robots fail in the field in ways the lab never shows. We're building the physical AI flywheel that lets your agents find and fix those failures first.

Every site is different and field days are scarce. Compute isn't. So the loop keeps going after your team goes home, and a better release is waiting in the morning.

  • Observe the physical worldField logs, telemetry and robot code start the loop.
  • Improve autonomouslyAgents close the loop between simulation and reality: more coverage and more fidelity, every night.
  • Release with confidenceThe whole suite runs on every change. Your team signs what ships.

How it works. From messy field data to a release you can sign.

Your dataas it is, nothing to reformatRobot: quad_x
ROS bags214 flightslidar, camera, IMU
Logs2,381 files3 formats
Config12 versionsdiffed
Robot code23 releaseslinked to flights
Site videos, scans3 sitesmatched to runs
Jira, Slack412 threads2 unclear
What it made of it
1 robot, 6 sensors, 3 sites14 flights since the last release2 incidents linked to tickets
Still unclear
  • Two bags have no timestamps
  • Which config flew on 12 Oct?
Ask the teamStart watching
Your data

Start from the data you already have

Logs, ROS bags, config, code, site videos, tickets. It's a messy world, and nobody has time to clean it up. Nonlimi makes sense of it as it is.

  • Nothing to reformat
  • Logs, code and tickets read together
  • Every fact points to its source
Generated scenariosfrom your flights and incidentsAdd 9 to the gate
ScenarioPass when
Pilot hands back to autopilot in a gustno jump above 0.5 m
Lidar drops out on the approachlands in a safe zone
GNSS lost under the bridge, 20 sdrift under 2 m
A person walks under the flight pathnever over a person
A second drone crosses the corridorkeeps 5 m apart
+ 4 more, each with its evaluators
9 proposed
3 safety-critical · 3 sensor · 3 actors
Left out
31 near-duplicates of cases your suite already covers
Scenarios

The scenarios that matter, generated for you

Writing scenarios by hand is the slow part, and the worst cases are the ones nobody thought of. Agents propose them, each with the evaluators that decide pass or fail.

  • Safety-critical moments
  • Sensor failures
  • People and other drones moving
What changedsince release 4.2Re-run 11 affected
Still standsNew from the fieldRe-run for this change
Config diff
lidar driver 2.1 → 2.3max speed 1.6 → 1.8 m/s
These touch 11 of 214 scenarios. The other 203 still stand.
Every release

Change the sensor, the config or the software. Keep your confidence.

Every release leaves scenarios and results behind, and the next run starts from them. Nonlimi re-runs only what a change touches and tells you which earlier results still stand.

  • Value stacks up, version by version
  • Config diff in plain terms
  • Only affected scenarios re-run
Release suite214 scenarios × 4 repeats · 3 hFailed
Position jump at handover (m)
4.24.3-rc1- - limit
Agent · 1 regression4 of 4 repeats agree, so the change caused it.Commit a91f3c changed the estimator gate. The position jumps at handover.
Test reportRe-run on the fix
Test runs

The whole suite on every change, in hours instead of weeks

Every build or parameter change runs your full suite in parallel, in a simulation kept true to your field data. You see what regressed and why, and get a report to sign before anyone drives to the field.

  • Hundreds of runs at once
  • Regressions explained from the logs
  • Go or no-go with a report
Why it failedbrick 1,204 placed 12 mm off6 layers checked
Where it could come fromEvidence
Camera calibrationreprojection error normalruled out
Localizationdrift under 1 mmruled out
Mechanical flexboom bends 9 mm under loadroot cause
Control timingno missed cyclesruled out
Planning logica patch would hide itnot the cause
Site modelscan matches the wallruled out
AgentA planner patch would pass the test and leave the cause in place. Check the boom mount, not the code.
Knowledge

Failures cross boundaries. We know where to look.

A misplaced brick could come from camera calibration, localization drift, mechanical flex, control timing, planning logic or an incorrect site model. An agent may propose a convincing software fix while the underlying cause remains.

  • Follows a failure across the stack
  • Rules causes out with evidence
  • Says when it isn't software

The goal isn't more testing. It's deployments that stick.

It is designed to compound: each release leaves fewer problems for your customers to find, and teams can release more often and at higher quality.

Incidents that come back↓ down
Field failure to redeploy↓ down
Field days per release↓ down
Releases that reach the fleet↑ up

Field tests go to what you are building now. We agree these four numbers with each team before we start.

Make your next deployment your best.

If you ship robot software, we'd like to hear from you.

Talk to us

Or follow along: our newsletter, when we ship something or publish research.

Talk to us

Tell us what you're working on. We'll get back to you.

Nonlimi Dynamics S.L. uses these details only to reply to you. To see, correct or delete them, write to info@nonlimi.com. Privacy policy

Thank you.

Your message is on its way to the team. We'll get back to you.