Cross the reality gap. On purpose.
A policy that wins in simulation and fails on hardware is worse than no policy at all. Kyros is built around the techniques that actually transfer: aggressive domain randomization, principled system identification, and ruthless transfer evaluation.
Domain Randomization
Mass, friction, motor dynamics, latency, lighting, textures, camera pose — every nuisance variable is parameterized and randomized per episode. Your policy has to be robust by construction.
System Identification
Calibrate your simulator against real-world rollouts. Kyros fits sim parameters to measured trajectories so your training distribution matches your robot.
Transfer Evaluation
Run any policy through a standardized eval suite that produces a transfer-likelihood score. Stop guessing whether your policy will work outside the lab.
Safety Sandboxing
First deployments happen in a hardware-in-the-loop sandbox with safety envelopes, action clipping, and automatic rollback on failure.
The numbers.
Average sim-to-real success rate across Transfer Verified environments
Improvement over default Isaac/MuJoCo randomization presets
From trained policy to first hardware rollout via the Kyros eval pipeline