WORLDTECHGoogle's GKE Labs puts reinforcement learning on Kubernetes with OpenRL
Google's GKE Labs group has turned model fine-tuning through reinforcement learning into a Kubernetes workload, and Google has donated the resulting project, OpenRL, to the Cloud Native Computing Foundation. SiliconANGLE reports that the approach separates the reinforcement learning loop from the compute plumbing beneath it.
What OpenRL provides
OpenRL debuted as a research preview in June 2026. It offers a self-hosted training API that a company can run on its own Kubernetes cluster, so an existing model can be further trained for a narrower job without the surrounding infrastructure having to be assembled from scratch.
Fine-tuning through reinforcement learning is becoming one of the more demanding jobs in AI infrastructure. It brings training, inference and sandboxed agent work into one loop, and the research code involved often sits tightly bound to the systems underneath.
How the GKE Labs team got there
According to Sunil Arora, a senior software engineer at Google Cloud, the team first built an AI agent for Kubernetes tasks and then a benchmark for grading models on that work. GKE Labs is the group that puts experimental projects into users' hands early.
Arora described a split in which infrastructure engineers handle scaling while researchers keep control of their experiments. Justin Santa Barbara, a developer at Google, also took part in the discussion.
Sandboxes and limited compute
Sandboxes are where the reinforcement learning loop meets agent infrastructure, because calculating rewards can mean running millions of isolated commands. Arora said Google's Agent Substrate allows more of these sandboxes to run on limited compute.
The discussion was recorded at KubeCon + CloudNativeCon and broadcast on theCUBE, and Google Cloud sponsored that segment.
Open questions
- Will OpenRL move beyond its research preview stage?