Real-world training
Average loss when lab policies meet real friction, contact noise, and hardware risk.
- Weeks of setup per task
- Robot damage and safety risk
- Impossible to scale variations
- No edge case coverage
Kyros lets every policy encounter scale, variation, and failure modes before the first hardware run.
Average loss when lab policies meet real friction, contact noise, and hardware risk.
Validated environments make policy training repeatable before a robot touches hardware.
# Kyros Platform SDK — v1.0import kyros# Load a pre-built kitchen environmentenv = kyros.environments.load("kitchen_v4.2", robot="franka_panda", variations=10000, domain_randomization=True)# Configure training curriculumcurriculum = kyros.curriculum.build( task="pick_and_place", difficulty="progressive", edge_cases=["clutter", "occlusion", "low_light"])# Train policy — runs 430,000x faster than real timepolicy = kyros.train(env, curriculum, timesteps=1e8)# Export for real-world deploymentpolicy.export("my_robot_policy.pt")Browse our marketplace of pre-built, physics-accurate simulation environments. Kitchen, warehouse, outdoor terrain, hospital room. Every major task category is covered.
Select a task from our library, set your domain randomization parameters, choose your curriculum difficulty progression. No environment-building required.
Export your trained policy. Our transfer-optimized environments are validated for real-world deployment. What trains in Kyros, transfers to your robot.
Hundreds of pre-built, task-specific environments. Kitchens, warehouses, outdoor terrain, hospitals. Ready in minutes, not months.
Pre-configured task definitions with reward functions, success criteria, and observation spaces. Don't write reward code. Just train.
Ordered difficulty progressions proven to produce generalizable policies. Start easy, end expert. Built by researchers who've done it before.
10,000 object variations per environment. Tested randomization configs that maximize transfer rate without breaking policy learning.
Built on Genesis — the physics engine that runs 430,000x faster than real time. Train policies in 26 seconds that used to take days.
Every Kyros environment is validated against real-hardware transfer benchmarks. Verified environments carry a 'Transfer-Tested' badge.
Kyros environments are productized simulation assets: geometry, physics, randomization, task definitions, and transfer reports move together.
Cabinets, transparent objects, soft clutter, lighting variation, and contact edge cases ship as one reproducible training environment.
Policies that look solved in simulation still fail when contact, friction, sensing, latency, and object variation change.
Most teams spend weeks building scenes, tuning randomization, and wiring reward logic before training even starts.
The winning stack needs reusable environments, reproducible runs, and transfer evaluation built into the workflow.
Join the waitlist. Be the first team to access Kyros and experience physics-accurate simulation at scale.