The safety layer for every robot release

Approved safety scenarios run on every change: replay first, then simulation, then real robots. Regressions get caught in the pull request, not on the floor.

Your policy changed on Tuesday.
Your safety tests were written in March.

MarchTuesday
Safety testsPolicy release

Robots learn now

Learned policies can behave differently after every release.

Testing is scarce

Manual, physical trials, and weeks to certify one machine.

The rules got stricter

ISO 10218:2025 and ANSI/A3 R15.08-3 now ask how you control software changes after a robot ships.

Evidence is one-time

A certificate covers one moment. The risk lives in every release.

From incident report to tested release

One incident becomes thousands of tested variations, each tied to the clause it would break.

fail · crouching · stopped at 0.41 m
01

Sources

Incident tickets, near-miss reports, robot logs, standards clauses and requirements.

A verdict on every pull request

Checks a certifier recognizes: separation distance, contact force, stopping distance, staying inside the workspace. Real regressions block the merge. Noise doesn't.

GitHub#1482

foretell / safety — regression found

#1482 nav: tighter cornering in narrow aisles

  • separation ≥ 1.0 m · blind corner2 failed
  • stopping distance · crouchingpassed
  • stays inside workspacepassed
  • pick rate · aisle 4−1.2%, within noise
  • coverage116 of 120 moments checked · 4 not checked (sensor gap), listed for review

People first.
Not only people.

  • People: too close, closing fast, appearing from behind a rack
  • Forklifts and other robots
  • Keep-out zones, dock edges, slow zones
  • Sensor and localization faults: does the robot slow or stop when a sensor drops out?
  • Arm speed and acceleration near people
  • Tip-over and contact force, in simulation and on real robots

One record.
Four readers.

Every release gets one record, exported for whoever needs it, with a log of every safety decision and sign-off: the change-control trail certifiers, insurers and your customers ask for.

release-2.14.record

What changed, which scenarios ran and why a change was stopped.

change
#1491 nav: wider margin at blind corners
compared with
release 2.13
scenarios run
1,240
result
passed · fixes the regression in #1482

Built for robot makers whose robots work near people

Teams of 10 to 100 engineers. Most start when:

  • They've just had an incident
  • A certification project is starting
  • An insurer asked for safety documentation
A mobile robot on a warehouse floor

Your simulator,
on your machines.

Nothing gets replaced. Your logs and robot software never leave your side; only results do. The dashboard runs in our cloud or yours.

foretell

Code & work

  • GitHub
  • GitLab
  • Jira
  • Linear

Robot logs

  • ROS 2
  • MCAP
  • ROS bags
  • LeRobot
  • Foxglove

Simulation

  • Isaac Sim
  • MuJoCo
  • Gazebo
  • Genesis

Log storage

  • Amazon S3
  • Google Cloud Storage
  • Azure Blob

Read-only connections.

We don't certify robots
TÜV, UL and exida do
We don't build safety hardware
Keep your scanners
We don't build simulators or digital humans
Bring yours

Questions

What does a green check mean?

No regressions found in the moments we checked, and it says how many that was. Anything we couldn't check, like a gap in the log, an uncertain detection or a moment where the new build drove differently, is marked "not checked", never "passed". Foretell never calls a robot safe; it tells you what was tested.

Do we need to log where people are?

No. Foretell's own detector finds people, forklifts and obstacles in your lidar, depth and camera data. If you do log your own tracking, we cross-check it, and any disagreement is marked "not checked".

Do we have to write every scenario ourselves?

No. An agent drafts scenarios from your incidents and logs, and you start from a library of common hazards mapped to standards. A safety engineer approves each one before it runs.

Does Foretell certify my robot?

No. TÜV, UL and exida certify. Foretell produces the release-by-release evidence you bring to them.

Do I need a simulator?

No. Replay of your own logs works on day one. Simulation is added for a type of scenario once it has proven accurate for it.

Can we trust scenarios an agent wrote?

Nothing joins your suite until a safety engineer approves it. Every scenario cites its source log and clause.

Does it help with the EU Machinery Regulation?

We're built around US standards first. Foretell keeps a record and a decision log for every release from day one, so if you sell in the EU once the regulation applies (January 20, 2027), the history is already there. A notified body still does the assessment.

Where does it run?

Everything that touches your logs, robot software and simulators runs on your machines. Only results (verdicts and numbers) leave. The dashboard runs in our cloud, or in yours if you need it.

Start with your last ten incidents

In 30 days they become approved scenarios that replay on every change, with your first release record. No simulator needed.