Skip to main content
Cadenza Labs builds the software infrastructure that turns a robot into a learning machine. We give physical-AI teams one stack to scaffold a robotics project, simulate it at scale, fine-tune a vision-language-action (VLA) policy on the rollouts, and deploy the result to real hardware — without stitching together a dozen disconnected tools. If you are training robots, you spend most of your time on plumbing: a simulator here, a data pipeline there, a training loop somewhere else, and a fragile bridge to the robot at the end. Cadenza collapses that into a single, coherent workflow.

​
What we do

Simulate physical AI

A MuJoCo-based physics core with batteries-included robots, scenes, and a reusable action library — so you can roll out missions in seconds, not weeks.

Generate training data

Every simulated mission is logged tick-by-tick and scored, producing clean rollouts ready to become fine-tuning data.

Fine-tune policies

Turn rollouts into VLA training sets and LoRA adapters with one command — reinforcement learning and supervised fine-tuning, without the boilerplate.

Deploy to hardware

Take the same policy from sim to a physical Go1 or G1 over DDS, SSH, or a bridge — the loop you tested is the loop that ships.

​
How the pieces fit

Cadenza is one stack with two layers. The SDK is the foundation — the physics, robots, and inference core you build on. The CLI is the workflow on top — it takes what you build and turns it into data, training, and deployments.
  • Cadenza SDK — the cadenza-lab Python library: the MuJoCo-based simulator, robots, action library, and inference stack. This is where you build — define scenes, drive a Gym loop, or plug in your own world model.
  • Cadenza CLI — the cadenza binary built on the SDK. This is where you operate — scaffold projects, gather rollout data, fine-tune policies, and deploy to hardware, from the command line or CI.
New here? Start with What we provide for the big picture, then build your first robot in the SDK quickstart.