Forge's AI turns plain English into a physically real twin of your site, robot and sensors. It drafts the scenarios and edge cases and runs the ones you approve, so failures show up in simulation, before deployment. Your autonomy stack runs as-is and your engineers decide what counts as a pass; Forge does the setup work that used to need a dedicated simulation team.
52 pass
2 fail
Fireloop Forge is a simulation-based evaluation platform for physical AI. It builds a physically calibrated digital twin of a customer's site, robot and sensors, drafts test scenarios and edge cases from plain-English descriptions, runs the robot's own autonomy software - unmodified - through every scenario the customer's engineers approve, and returns a graded, replayable verdict before deployment. Fireloop, the company behind Forge, builds on NVIDIA Isaac Sim and Omniverse and is an NVIDIA Inception member. Forge is built for mobile robots that move goods - autonomous forklifts, pallet movers, tuggers and goods-to-person AMRs - with the forklift as the worked example on this site.
The problem
A leaning pallet, a person stepping into the path, light that washes out the camera. Today these moments turn up one at a time, months into a field pilot, with people nearby.
Robot arrives on siteDemo in a cleared area goes perfectly
Stops short in morning glarelightingCamera washed out · shift delayed
Near-miss with a person stepping outpeopleFleet paused pending safety review
Unexpected obstacle blocks the routeobstaclesRobot halts · operations stall
Software update undoes the Month 3 fixregressionNobody re-tested. Nobody noticed.
Every edge case waits for the real world to produce it.
Scenes, robot models, sensor calibration and test code all need scarce skills.
Each variation is built by hand, so most are never tried.
What worked last quarter quietly breaks in the field.
One platform
Develop, evaluate, and generate data in one place - anchored on the piece the others exist to serve: an independent verdict you can stand behind.
Harden your driving policy against a calibrated digital twin of the site it will actually deploy into.
Prove function, safety, and benchmark in one loop - a graded, standards-traceable result for your policy as shipped, ready to show a buyer.
Turn every run into labeled, ground-truth scenario data - the long-tail coverage a real floor can't give you.
// 01 Forge builds your world
Describe the site or import a video. Import your robot or pick an existing one. Forge's AI lays out the scene, sets up the robot, then places and calibrates its sensors, so the robot moves and sees in simulation the way it does in real life. You review each step and correct anything Forge's AI got wrong.
Building from the asset library…
Calibrating to the real sensors
// 02 Forge writes your tests, runs and debugs them
In the world you just built, describe the situation in plain English. Forge's AI turns it into a runnable scenario, widens it with the conditions and edge cases the real world throws at robots, and runs the set you approve.
AI drafted the scenario
AI suggested edge cases for this scenario
Stopped at 0.28 m. Needed 0.50 m. Fail.
7 pass · 2 fail, found before anyone was at risk.
/camera/front exposure saturated for 340 ms after the door opened → detection came 0.6 s late → braking started at 1.1 m instead of 1.7 m.
Every run keeps the robot, site and settings that produced it, ready to replay for your safety review.
robot v3.2 · site DC-3 · seed 4127// 03 Why Forge
Forge handles the scene, robot, sensors and scenario drafting. Your engineers review, approve and decide what counts.
Calibrated robot movement and sensors in a physics-based twin, so what happens in Forge is a faithful predictor of what happens on site.
When a detail is unclear, Forge's AI asks. When a change would alter what the test means, it says so.
Every result carries the measurement behind it and replays deterministically, ready for your safety review.
Forge never modifies, retrains or wraps your policy. The same build that drives in the field drives in the twin, so a pass means your release passes.
// 04 Solutions
The workflow is the same either way: describe the world, let Forge draft the world and the tests, approve them, read the verdict. What changes is whose question it answers.
Forge takes the grunt work out of simulation - scenes, assets, sensor models, scenario scaffolding - and hands your engineers a root cause instead of a symptom. Your stack runs unmodified.
Forge for OEMsTest robots from any vendor against your site, your processes and your business case - before go-live and after every update. For deployers, OEM sales teams and system integrators.
Forge for deployers// 04 Who it's for
Will this release hold up at every customer's site?
Recreate each customer's site with Forge, run every release against the situations they actually face, and catch what broke before they do.
38 of 40 hold · 1 fixed · 1 new failure
Can I show this prospect it works in their building?
Rebuild the prospect's site from a description or a walkthrough video, run your robot through their busiest hour, and walk into the pitch with a graded result for their floor instead of a generic demo.
12 of 12 ready · 1 fixed before the pitch
Will this robot actually work at my site?
Describe your site and how it runs. Forge builds it and tests the robot there before go-live, with your team reviewing every scenario. No code required.
11 of 12 ready · 1 needs attention
// 05 Questions
The things teams ask us first: what Forge does to your stack, what stays inside your network, and how it differs from the simulator you already run.
Fireloop Forge is a simulation platform that evaluates a robot's autonomy software against a physically calibrated digital twin of the site where it will deploy. Forge's AI drafts the scene, robot setup, sensors and test scenarios from plain English; your engineers review and approve each step; then Forge runs your unmodified policy through every approved scenario and reports pass or fail with the measurement that decided it.
No. Forge's AI does the work that used to require a dedicated simulation team: laying out the scene from a description or walkthrough video, importing and tuning the robot model, placing and calibrating sensors, and drafting scenarios. Your team's job is to describe the site, review what the AI proposes, and decide what counts as a pass.
Two audiences. Robot OEMs, to regression-test each release against real customer sites, hand their engineers a root cause instead of a symptom, and walk into pilots with evidence. And the people putting robots to work - warehouse and logistics operators, OEM sales and field teams, and system integrators - to test any vendor's robot against their own site, processes and business case before go-live and after every update.
Book a 30-minute demo. Bring one scenario that worries you; we build your world and run it in Forge with you during the session.
You describe the site in a sentence or import a walkthrough video. Forge's AI lays out racking, docks, aisles and zones from sim-ready assets, names the areas so scenarios can refer to them, and flags anything it is unsure about for you to confirm.
You state what should happen in plain English, for example "the robot must stop at least 0.5 m from a person stepping out in Aisle A-7." Forge's AI drafts each step and assertion, asks a question when a detail is ambiguous instead of guessing, and suggests edge cases and condition sweeps - speeds, lighting, sensor noise, occlusion - that you can accept or reject.
Forge reports the measurement that decided the result (for example, stopped at 0.28 m against a 0.50 m limit), keeps the exact robot, site and settings so the run can be replayed deterministically, examines the ROS messages to propose a root cause, and reports failures that recur across runs or releases as trends.
No. Forge models the vehicle's physics and sensors; your driving or navigation policy connects over ROS and makes every decision itself. The same build that drives in the field drives in the twin, so a pass in Forge is a pass for your release, not for a simulation-friendly variant.
No. Forge evaluates the policy as a black box over a secure link. Model weights, source code and proprietary logic stay inside your environment; Forge only exchanges the sensor inputs and control outputs needed to run the scenario.
Isaac Sim and Gazebo are simulation engines. Forge is an evaluation layer that runs on top of Isaac Sim: it builds the site-specific twin, authors and varies the scenarios, connects your policy unmodified, grades each run against limits, and produces replayable evidence. Teams typically spend months building that layer themselves; Forge provides it.
A field pilot surfaces edge cases one at a time, over months, with people nearby. Forge reproduces the same conditions by the thousand before the robot reaches the floor, and re-runs them on every software release so regressions are caught before a customer sees them. Forge complements pilots; it does not replace the final on-site validation.
Forge is built for mobile robots that move goods: autonomous forklifts, pallet movers, tuggers and goods-to-person AMRs. The forklift is the worked example on this site because it's where the stakes are highest; for other mobile robots the drive model and sensor suite change and the workflow doesn't. Humanoids, drones and field robots are not a focus today.
The forklift suites reference ISO 3691-4 and ANSI B56.5 for safety scenarios and VDA 5050 for fleet interoperability. Results are traceable to the requirement each scenario was written against.
Book a demo
We'll build your world and run it in Forge with you in a 30-minute session.