The science behind Kyros.
Our platform is built on the latest research in sim-to-real transfer, domain randomization, and physics simulation. We share what we learn.
Papers we've built on.
The academic foundation of the Kyros platform.
DROPO: Sim-to-Real Transfer with Offline Domain Randomization
A novel approach to narrowing the sim-to-real gap through posterior estimation of physics parameters from offline real-world data.
Humanoid-Gym: RL for Humanoid Robot with Zero-Shot Sim2Real Transfer
Demonstrating zero-shot sim-to-real transfer for humanoid locomotion — trained entirely in simulation, deployed without real-world training data.
AutoMate: Specialist and Generalist Assembly Policies over Diverse Geometries
Generalist assembly policies trained over thousands of geometries. Sets the standard for contact-rich manipulation in simulation.
A Survey on Sim-to-Real Transfer Methods for Robotic Manipulation
A comprehensive survey of state-of-the-art techniques for closing the sim-to-real gap in robotic manipulation.
Our technical foundation.
The most advanced open-source physics simulator ever built. 430,000x real-time speed. Apache 2.0 licensed.
Industry-standard contact dynamics. Battle-tested in DeepMind research. The gold standard for manipulation physics.
GPU-accelerated robot learning. Used by Figure AI, Agility Robotics, and the world's leading humanoid companies.
The foundation of modern robot learning. All Kyros training pipelines are PyTorch-native.
The standard for robot deployment. Kyros environments export ROS 2 compatible policies.
800,000+ 3D objects for procedural environment generation. The content library behind Kyros's variation engine.