Simulation Infrastructure for Physical AI
Trekion rebuilds real facilities as photoreal and physics-accurate simulation environments and assets, so you can train and evaluate your robots before they reach the actual site.
Real-world data collection and policy evaluation cannot scale.
Robot data is scarce
Robots need task-specific data, but collecting it requires site access, supervision, and engineering time. That makes it expensive to scale.
Off-the-shelf environments and assets fall short.
Off-the-shelf environments and assets rarely match your site or its physics. A simulation you cannot trust is not worth building.
Physical testing is costly and hard to repeat.
Testing requires hardware, people, and access to the right space. Some tests are impossible to run.
Real world facility to a working simulation.
We turn facility walkthroughs and product photos into simulation-ready environments and assets. On the left is what the real camera captured. On the right is the scene recreated in simulation, complete with collision meshes. We aim to match the deployment environment as closely as possible, so you can test before the robot reaches the site.
Products.
Environment Creation
Your site, or any space, rebuilt as a photoreal and physics-accurate scene you can train and test in.
Explore environment creation
SimReady Assets
One photo in, a physics-ready asset out. Mass, friction, collision mesh, exported to USD, MJCF and URDF.
Explore SimReady assetsCurrent success rate
86%
Latest checkpoint
SmolVLA Baseline
Best success rate
100% pass
Evaluation
In developmentRun your policies against scenarios built from your own site.
Talk to usFrom a real building to a working simulation.
Capture
A walkthrough of the real space: video, photos or scans. Works from CAD designs too
Reconstruct
We reconstruct the capture into a simulation that is photoreal for a camera and accurate for a solver, with measured physical properties on the objects in it.
Use
Train and evaluate in it, generate variations, and export to Isaac Sim, MuJoCo or Gazebo through OpenUSD.
Tell us what you are building.
A 30-minute technical call. We will map the pipeline to your stack and put together a sample scene from your spec.
- 30-minute technical call
- Straight to the founders
- A sample scene from your use case
Common questions.
How is this different from benchmark evaluation?
Benchmark platforms score your policy on standard task suites. We rebuild the facility it will actually ship to: that site's racking, floors, lighting, and traffic.
Why does site-specific matter? Is generic sim not enough?
Generic simulators reproduce the task but not your building. They do not have your racking, floors, label formats, lighting, or the traffic a fleet actually meets. Work on real deployments keeps showing that grounding a simulation in the specific site is what makes the results worth acting on. Generic gets you part of the way there. The site is the rest.
Which stacks does it work with?
Scenes and assets export to OpenUSD, Isaac Sim, MuJoCo, and Gazebo.
Do you need access to my facility?
For a site twin, yes: scans, video, or floor plans. For generated environments and asset packs, you can start without any facility access.