Red Hat and HP: Empowering physical AI from bench to fleet

Red Hat and HP: Empowering physical AI from bench to fleet

Component | Article_teaser

October 21, 2026

4

-minute read

Artificial intelligence

,

Edge computing

Jeremy Ary

Senior Principal Software Engineer

Subpattern | social_share_set

Component | social_share

Share

Subpattern | subscribe

Component | social_icon

Subscribe to RSS

Component | Icon

© Red Hat, Inc. CC-BY-4.0 licensed

Subpattern | results_nav

Group layout

Component | Nav_links

• Back to all posts

Component | Generic

Teaching a robot to perform its first task is thrilling, but turning that initial win into a repeatable, self-improving process capable of backing production fleets is a challenge. Communities like the one behind Robot Operating System 2 (ROS2) have made the record-train-deploy loop more accessible than ever, but challenges remain:How can we assess policy quality across iterations?

• Does the policy improve or is it a fluke based on a series of fortunate episodes?

• Can we quantify good and bad episodes contributing to successes and failures?

• Can we trace lineage and verify the authenticity of our artifacts?

• Can we properly simulate, scale, and manage fleets?

We can already answer all of these questions, but doing so often requires a variety of tools across a multitude of systems following manually coordinated processes. Given the continued, rapid growth of physical AI, these approaches often become outdated and unmanageable fairly quickly. Red Hat and HP have joined efforts to offer a more grounded approach to tackling these challenges.

Visible policy improvement

We start with the ROS community’s own physical AI demo , leveraging Gazebo to simulate an open source SO-ARM101 arm using a policy trained with LeRobot to pick up and place 3 cubes on a tray. As a challenge, we depart from the exact environment used for training the original policy by randomly shifting 1 cube up to 3 cm in any direction, thus introducing a need for new behavior to account for the variance of placement. As we run episodes of the established policy performing its task against the new circumstances, we evaluate outcomes and categorize successes and failures. We collect successful runs into versioned training data and note failures for traceability, but omit them from collection.

Custom flywheel dashboard showing robotic arm in action and episodic evaluations

After the curated training dataset reaches a configurable threshold event, a Red Hat OpenShift AI pipeline trains a new policy candidate using the collected successes, evaluates its performance against the existing policy on identical scenes, and, when it finds meaningful improvement, packages it into a signed container and registers it with the model registry. A pull request created by the pipeline gates promotion of the candidate to provide an opportunity for human review. Following promotion, Red Hat Edge Manager verifies and delivers the model, and the loop kicks off with data collection once more.

Early candidates trained on too little data became worse by comparison against the original policy and failed to pass evaluation gates, proving regression is both possible and impactful. When trained with an ample dataset, however, our resultant policy climbed from the 81.9% success rate of the original policy to 92.5% across 360 identically seeded scenes.

Data center AI power at the desktop

Observability graphs indicating compute and memory use across the 4 tenants

Everything involved with running the loop and simulating large fleets resides on 1 machine. The simulations, training pipelines, evaluation gates, platform governance, and managed devices all run on a single HP ZGX Fury AI station. Backed by NVIDIA’s GB300 Grace Blackwell architecture and 748 gigabytes (GB) of coherent memory, the ZGX Fury handles our workload using a fraction of its available resources, leaving substantial headroom for tasks like reinforcement learning (RL)-based training across numerous parallel simulations, hosting large-scale models up to 1 trillion parameters, or running extensive digital twin environments, all alongside core development efforts supporting multi-tenant teams.

HP provides a powerful system backed by NVIDIA accelerated architecture to enable more workflow stages to run locally, and Red Hat provides continuity starting with data and evaluations running through to trust management, delivery, and recovery. Teams no longer need a separate cluster for training, nor a secondary environment for validation. There's no need to wait for data center assistance and reporting of outcomes. No queuing for regional resource availability in public clouds, transmitting proprietary data outside of your own network, or dealing with large amounts of egress bandwidth. From development to production, it's all at the workstation.

To further strengthen our offering and collaboration, HP shared a plan to develop an open enterprise AI platform designed to run production inference closer to users, applications, machines, and data. The planned solution will combine the HP ZGX Fury with Red Hat AI Factory with NVIDIA , enabling enhanced AI and orchestration capabilities native to NVIDIA’s accelerated architecture. The work performed to enable our flywheel loop here serves as an exploration of similar concepts through robotics, while the platform aims for much more, like helping companies maximize local AI inference throughput, reduce environment setup time and deployment risk, improve GPU utilization through optimized CUDA libraries, scheduling, and workload orchestration, and manage AI models, agents, and applications across hybrid cloud.

Empowering the robotics developer

Our goal for the developer is to keep the tools and methods you already know and use, but lose the coordination work necessary to tie them together. Keep the training and simulation in established tools, but let the platform assist with evaluation, artifact signing, and fleet management. Simplify to a single AI station rather than deal with a litany of deployment environments, configurations, and clouds. Get readily available answers on lineage and statuses without referencing complicated spreadsheets or tracking by memory. Produce trusted, traceable artifacts through pipeline automation rather than manual wrap-up processes at the end.

Containers yield continuity across environments, from laptop to workstation and data center, and we want to help developers embrace that. In the project’s loop, our pipeline produces a signed Open Container Initiative (OCI) policy image. We don't repackage anything between the candidate policy passing evaluation and the model loading into the device. To support that workflow extending back to the developer’s environment, Red Hat is building an open source physical AI extension for Podman Desktop that builds images from a chosen ROS2 distribution and architecture, runs and shows simulations, monitors ROS2 topics live, aids with registry work, cluster deployment, and more. We intend this to be an upstream project that can complement whatever tooling a workstation ships with, and we welcome hardware partner and community member contributions.

Podman Desktop robotics extension

Ready to streamline your physical AI development from bench to fleet?

Discover how Red Hat powers business-critical industrial systems at redhat.com/industrial , and read HP’s recent press release to see how we're extending enterprise AI architecture to the operational edge.

Block | Dynamic pattern deluxe promo

Component | Band_header

Resource

The adaptable enterprise: Why AI readiness is disruption readiness

This e-book, written by Michael Ferris, Red Hat COO and CSO, navigates the pace of change and technological disruption with AI that faces IT leaders today.

Component | spacer

Component | Cta_multi_basic

Subpattern | simple_cta

Component | CTA

Get the resource

Deluxe mbox

Component | Card_header

About the author

Subpattern | speaker

Card layout

Component | Image_embed

Component | Person

Jeremy Ary

More like this

Dynamic pattern

Blog post

RHEL for NVIDIA delivers trusted foundation for next-gen AI workloads

Securing AI agents in the enterprise AI factory: NVIDIA BlueField and DOCA Argus on Red Hat OpenShift

Original podcast

Can agentic AI patch the internet?

How Red Hat cleared IT debt for scalable AI

Subpattern | card_flex

Subpattern | text_basic

Keep exploring

• What is agentic AI? Article

• Predictive AI vs. generative AI Article

• Top considerations for building a production-ready AI/ML environment E-book

• Generative AI, the Ansible way Video

• Innovate and transform with a modern application platform E-book

Keep Exploring mbox

Subpattern | simple_text

Browse by channel

Explore all channels

Pattern | raw_html

Automation

The latest on IT automation for tech, teams, and environments

Artificial intelligence

Updates on the platforms that free customers to run AI workloads anywhere

Open hybrid cloud

Explore how we build a more flexible future with hybrid cloud

Security

The latest on how we reduce risks across environments and technologies

Edge computing

Updates on the platforms that simplify operations at the edge

Infrastructure

The latest on the world’s leading enterprise Linux platform

Applications

Inside our solutions to the toughest application challenges

Virtualization

The future of enterprise virtualization for your workloads on-premise or across clouds