Google gives OpenRL training software to an open source group

Google gave OpenRL to the Cloud Native Computing Foundation. OpenRL trains AI models with rewards on a Kubernetes cluster of a company. Kubernetes is software that controls groups of computers. Sunil Arora, a senior software engineer at Google Cloud, told about it at KubeCon + CloudNativeCon North America. OpenRL first came out as a research preview in June 2026.

How it works

Reinforcement learning trains a model with rewards. It mixes training, sampling and agent work in one loop. OpenRL keeps that loop apart from the work of the computers. Infrastructure engineers do the scaling, and researchers keep control of their tests, Arora tells.

Rewards can use millions of isolated commands. Google tells that its Agent Substrate puts more sandboxes on some computers. A sandbox is a closed space where code cannot cause damage external to it.

The proof

Google gave no test results, no speed numbers and no cost numbers.

What is not known

The source does not tell who uses OpenRL at this time. Justin Santa Barbara, a Google developer, told that the kubectl-ai tool could in theory format a hard disk. Santa Barbara wanted a sandbox for that risk.

Insight: Google gives away the software and has the chips

On 10 November, Google Cloud showed an open stack of six parts that includes OpenRL. It also tells that A5X servers, made for reinforcement learning, open to all customers in November. Open software on Kubernetes reads as a plan to move training work to Google Cloud chips. Monitor if other clouds use OpenRL. If they do not, it is not neutral.