The software infrastructure your robots can depend on.
Build, train, and deploy software for your robots in under an hour. Ship robots quickly, get deployment-readiness that won't disappoint.
Measured against current RL pipelines
Cost-effective. GPU-efficient. Deployment-ready.
Time to fine-tune · vs. training data
The more data you train on, the further Cadenza pulls ahead, fine-tuning in a fraction of the time current RL systems need.
01 · Infrastructure
One reliability layer, training to field.
Cadenza isn't a single tool. It's the infrastructure that makes physical AI deployment-ready, with four building blocks that span the whole lifecycle, on-training and in-field.
RL environments
Novel, cost-effective, GPU-efficient environments that auto-calibrate physics against real telemetry, so policies cross the reality gap on the first deploy.
Data generation
Generate domain-randomized training data at scale across hundreds of thousands of parallel instances. Reliable coverage of the cases your robot will actually hit.
Model evals
Benchmark deployment-readiness before a policy touches hardware. Seeded, reproducible evals that tell you what will hold up in the field.
On-board RL tokens
RL tokens that run on-board your robots after deployment, keeping policies adapting and reliable long after they ship.
02 · The SDK
Plug into the reliability layer in a dozen lines.
The SDK is one way in, not the whole product. It's how devs reach the infrastructure: name the robot and the scene, and Cadenza builds the environment, auto-calibrates physics against real telemetry, and scales to hundreds of thousands of parallel instances. The same G1 you train is the one you drive below.
- Declarative tasks. Reward functions defined inline and versioned, not glued code.
- Sim-to-real fidelity. Domain randomization tuned to your captured robot telemetry.
- Linear scaling. 100k deterministic instances per GPU partition, one seed.
1import cadenza as cz2 3# Build a MuJoCo environment for the Unitree G1 on Cadenza Lab.4# Cadenza auto-calibrates physics against captured telemetry.5env = cz.Environment(6 robot=cz.robots.UnitreeG1(dof=29),7 backend="mujoco",8 scene="lab/flat_floor",9 fidelity="sim2real", # domain-randomized, contact-rich10)11 12@env.task("walk")13def reward(state, action):14 speed = cz.metrics.base_velocity(state, axis="x")15 upright = cz.metrics.torso_upright(state)16 effort = cz.metrics.energy(action)17 return 3.0 * speed + 1.5 * upright - 0.01 * effort18 19# Scale to 100k parallel instances on one GPU partition.20fleet = env.parallelize(n=100_000, seed=7)21 22# Train a whole-body policy. Cadenza streams rollouts at 500 Hz.23policy = cz.train(24 fleet,25 algo=cz.algos.PPO(clip=0.2, gae_lambda=0.95),26 steps=2_000_000,27)28 29policy.export("g1_walk.cadenza") # deploy-ready checkpoint03 · In the field
Reliability that ships with the robot.
Deployment-readiness doesn't stop at training. Give the Unitree G1 a goal and on-board RL tokens keep perception, orchestration, and the learned policy adapting in closed loop. It's the same reliability layer, now running in the field.
CadenzaPerception
proprio + terrain → state
Conductor
goal → gait / skill plan
g1_walk.cadenza
29-DoF whole-body @ 500 Hz
1import cadenza as cz2 3env = cz.make("cadenza-lab/g1-locomotion") # Unitree G1 in MuJoCo4policy = cz.policy.load("g1_walk.cadenza") # 29-DoF whole-body5 6obs = env.reset(goal=PROMPT)7while not env.done:8 action = policy(obs) # goal-conditioned9 obs = env.step(action) # MuJoCo @ 500 Hz10 env.render() # stream qpos to viewer04 · Who it's for
Reliability infrastructure for both sides of physical AI.
For inference startups
A rebuilt rollout engine reclaims 70% more GPU headroom, so you can serve more physical AI workloads per partition without buying more silicon.
For robotics developers
Everything you need to take a policy from sim to a shipping robot: environments, data, evals, and on-board RL tokens in one layer.
GPU-efficient by design
Cost-effective rollouts that scale linearly to 100k deterministic instances per GPU partition, one seed.
Reliable across the gap
Contact-rich physics auto-calibrated against captured telemetry, so policies cross the reality gap on the first deploy.
Any robot description
URDF, MJCF, or a Cadenza robot spec. Bring an arm, a hand, or a quadruped, and the infrastructure adapts to your hardware.
Deployment-ready, end to end
Export a single artifact that runs in sim and on hardware, with on-board RL tokens keeping it reliable in the field.
Make your physical AI
deployment-ready.
Cadenza Labs is the reliability layer for the next generation of physical AI, in training and in the field. Request access and we'll stand up the infrastructure with you.