KYROS
API Documentation

A Pythonic API your ML team already knows how to use.

Gym-compatible environments. Vectorized rollouts. First-class support for PyTorch and JAX. Hosted training infrastructure that scales from one GPU to a thousand.

Python 3.11
import kyros
# Load a Transfer Verified environment
env = kyros.make("kitchen_residential_v4.2", randomize=True)
# Train your policy with the standard PPO loop
trainer = kyros.PPOTrainer(env, num_envs=4096)
policy = trainer.fit(steps=10_000_000)
# Score it against the transfer eval suite
report = kyros.evaluate(policy, suite="manipulation_v1")
print(f"transfer_likelihood: {report.score:.2f}")

Five lines from import to evaluation.

Every Kyros environment implements the standard Gymnasium interface, returns batched observations, and integrates with the trainer of your choice.

  • • Gymnasium-compatible environments and wrappers
  • • Vectorized: 4,096+ parallel envs on a single GPU
  • • Reference trainers (PPO, SAC, BC) included
  • • Bring your own trainer — it just works
  • • Deterministic seeding and full run reproducibility

SDK surface area.

Environments

make(), list(), versions(). Load, fork, and version any environment from the marketplace.

Trainers

PPOTrainer, SACTrainer, BCTrainer. Or pass any callable as a custom trainer.

Hosted runs

kyros.cloud.launch() to run a training job on managed GPU infrastructure with full logs.

Eval & webhooks

Programmatic transfer evaluation, with webhook callbacks when policies cross thresholds.

Get the SDK. Start training.