Test every change
before it reaches a robot.
Invariant is test and evaluation infrastructure for robots and other physical AI. Run your autonomy stack, perception model or manipulation policy through repeatable scenarios, score it against ground truth, and see exactly what changed. Built for teams shipping faster than they can validate.
While you develop
Run any scenario from your laptop and see the result in minutes.
On every commit
Gate merges on the metrics your team cares about.
Nightly, at scale
Run your full suite of scenarios and trials in parallel.
Before release
Sign off on a build with evidence, not a gut feeling.
Every autonomy build,
tested before it ships.
Drive the same routes in simulation or replay recorded logs, from your laptop or in CI. Invariant scores pose error, clearance and goals reached against ground truth, and compares them with your last good build.
Catch regressions
before the field does.
Run your detector on the same labeled scenes, simulated or recorded, while you train and on every commit. Every detection is scored against ground truth, and a scene that passes on main but fails on your branch blocks the merge.
Grade every policy
before it grabs anything.
Evaluate every checkpoint on the same pick-and-place tasks, hundreds of trials in simulation, scored per object against the last one. A better average can't hide a new failure.
One harness, from first commit to release.
Connect your stack
Point Invariant at your autonomy stack, model or policy. Works with ROS 2 and rosbag2 (MCAP) recordings.
Run scenarios
Run the same scenarios, simulated or replayed from logs, with ground truth for pose, objects, paths and task success.
Compare and decide
See what got better or worse against any baseline, and stop regressions before they reach a robot.
Looking for a few teams to pilot with.
We're building Invariant with a small group of design partners across autonomy, perception and manipulation. If testing can't keep up with how fast you ship, let's talk.
or email founders@invariant.run