Red Hat and HP Show a Self-Improving Robot Policy Loop on One AI Station

A flywheel built on the ROS community demo

The demonstration starts with the ROS community's own physical AI demo. It uses Gazebo to simulate an open source SO-ARM101 arm, running a policy trained with LeRobot to pick up and place 3 cubes on a tray. To test the policy, the team moved 1 cube by up to 3 cm in any direction, which created a need for new behavior.

Episodes of the existing policy were run against the changed setup, and outcomes were sorted into successes and failures. Successful runs go into versioned training data. Failures are noted for traceability but left out of the collection.

Once the dataset reaches a configurable threshold, a Red Hat OpenShift AI pipeline trains a new policy candidate and compares it with the existing policy on identical scenes. If it finds meaningful improvement, the pipeline packages the candidate into a signed container and registers it with the model registry. A pull request created by the pipeline gates promotion, so a human can review it. After promotion, Red Hat Edge Manager verifies and delivers the model, and data collection begins again.

Results, including a failed regression

Red Hat reports that early candidates trained on too little data performed worse than the original policy and did not pass the evaluation gates. The company says this shows regression is both possible and impactful. With an ample dataset, the resulting policy climbed from the original's 81.9% success rate to 92.5% across 360 identically seeded scenes.

The post frames the loop as an answer to questions that robotics teams face: whether a policy truly improves or just had lucky episodes, how to trace lineage and verify artifacts, and how to simulate and manage fleets. Red Hat says these can already be answered, but usually with many tools and manually coordinated processes.

One workstation, and a planned platform

Simulations, training pipelines, evaluation gates, platform governance and managed devices all run on one HP ZGX Fury AI station. It is built on NVIDIA's GB300 Grace Blackwell architecture with 748 gigabytes of coherent memory. According to Red Hat, the workload used a fraction of the machine's resources. That leaves headroom for reinforcement learning across many parallel simulations, hosting models of up to 1 trillion parameters, or running extensive digital twin environments.

Red Hat argues that teams then need no separate training cluster or validation environment, no queuing for public cloud capacity, and no transmission of proprietary data outside their own network.

HP also shared a plan for an open enterprise AI platform to run production inference closer to users, applications, machines and data. It would combine the HP ZGX Fury with Red Hat AI Factory with NVIDIA. The stated aims include higher local inference throughput, shorter environment setup time, lower deployment risk, better GPU utilization and management of models, agents and applications across hybrid cloud. Red Hat describes the robotics loop as an exploration of similar concepts.

What it means for robotics developers

Red Hat says developers should keep their familiar training and simulation tools while the platform handles evaluation, artifact signing and fleet management. The pipeline produces a signed Open Container Initiative (OCI) policy image, and nothing is repackaged between a candidate passing evaluation and the model loading onto the device.

Red Hat is also building an open source physical AI extension for Podman Desktop. It is meant to build images from a chosen ROS2 distribution and architecture, run and show simulations, monitor ROS2 topics live, and help with registry work and cluster deployment. Red Hat intends it as an upstream project and invites contributions from hardware partners and community members.