Build the cell in software first

We model your robot, parts, cameras and cell in the simulator that suits the job. The same world then produces labeled training data and runs the tests every model has to pass before it touches hardware.

Start a project How we choose models

The same 239,361 learning actions in a simple simulated world, against a real arm at 3 seconds an action.: 8 min vs 199 hThe same 239,361 learning actions in a simple simulated world, against a real arm at 3 seconds an action.
MuJoCo physics steps per second for a humanoid on a laptop processor. Not every job needs a data-center GPU.: 650,000MuJoCo physics steps per second for a humanoid on a laptop processor. Not every job needs a data-center GPU.
Minimum video memory for Isaac Sim, on an RTX-class card. It does not run on macOS, which shapes the toolchain.: 16 GBMinimum video memory for Isaac Sim, on an RTX-class card. It does not run on macOS, which shapes the toolchain.

What we do

Simulation setup and worlds
Your robot, cell, parts and sensors modeled to match reality.
- Gazebo, MuJoCo, NVIDIA Isaac Sim and Isaac Lab
- Robot models from URDF, MJCF or USD
- Cameras, depth and force sensors that match real hardware
- Friction and contact tuned against measurements from the real cell
Synthetic data and model evaluation
Labeled data and repeatable test scenes from the same simulation.
- Randomized lighting, textures, poses and clutter
- Labels for detection, segmentation, pose and grasps
- Benchmark scenes with success-rate reports
- Sim-to-real checks against real camera images

Model or no model

Can the simulation be trusted for this decision?

Programmed approach

Hand-built scenes with measured friction and masses
Fixed test layouts scored the same way every run

Learned approach

Learned dynamics or world models trained on simulated pushes
Generated scene variations for data and evaluation

How we choose

Appearance and geometry transfer well. Contact does not yet: no published work claims to have closed that gap. So we use simulation to choose and rank approaches, then always confirm on hardware.

Choosing a simulator

No simulator is best at everything. We pick per project, and sometimes use two.

Simulator	Best at	Watch out for
Gazebo (Jetty LTS)	ROS 2 integration, camera work; supported to 2031	Slow for large-scale data generation and learning
MuJoCo	Fast, accurate contact; runs well on a laptop	Photoreal rendering is limited
NVIDIA Isaac Sim / Isaac Lab	Photoreal rendering and massively parallel RL on GPU	Needs an RTX GPU; parts of the stack are closed