# Igniting a Robotics Revolution

**Customer:** Physical Intelligence

Physical Intelligence has revolutionized its AI training on WEKA® NeuralMesh™, cutting data storage costs by 80% and delivering seamless hybrid, multi-cloud support.

## Results

- **10-15%** Faster Model Checkpoint Times
- **80%** Reduction in Data Infrastructure Costs
- **100%** Data Portability Across Diverse Deployment Types

Introduction

## Pioneering the Integration of General-Purpose AI in the Physical World

[Physical Intelligence](https://www.pi.website/) is developing advanced foundation models and learning algorithms to create a unified robot intelligence capable of powering everything from everyday tools to complex, industry-grade robots. Their groundbreaking general-purpose robot foundation model, π0, is designed to achieve dexterous task performance at an unprecedented scale by training on a vast and diverse set of internal and external data sources.

The Challenge

## Slow Model Checkpoints and Rigid Data Management

Physical Intelligence's AI model training environment demands high performance, efficient data management, and robust scalability to power its groundbreaking robotics applications. Training its general-purpose robot foundation model required vast, diverse datasets and dexterous task performance running on Oracle Cloud Infrastructure (OCI). However, Physical Intelligence grappled with slow model checkpoints and rigid data management, which resulted in limited efficiency and scalability.

### Long Model Checkpoint Times

Traditional data storage on OCI resulted in prolonged checkpoint operations that delayed model training by 10–15%, hindering rapid innovation.

### Rigid Multi-Cloud Data Management

As the team sought a data management solution capable of operating seamlessly across their current multi-cloud environment—and future on-premises deployments—the limitations of conventional storage solutions became evident.

> Since we are a lean team dedicated to research, we needed a data storage solution that works out of the box for large-scale AI training and data management—handling the heavy lifting and the small details for us so we can stay focused and move fast.
>
> — Ury Zhilinsky, Member of Technical Staff • Physical Intelligence

The Solution

## A Petabyte-Scale Single Namespace

Physical Intelligence turned to the NeuralMesh to overcome these obstacles. With WEKA's Converged Mode for OCI, the team implemented a petabyte-scale single namespace that spans both high-performance NVMe-based storage and Oracle Cloud Object Storage capacity tiers.

> WEKA turns unused local NVMes into a high-performance shared storage system that is resilient and available.
>
> — Ury Zhilinsky, Member of Technical Staff • Physical Intelligence

The Outcomes

## Slashing Model Checkpoint Times

Physical Intelligence's model checkpoint times dropped significantly, boosting overall training speed by 10–15%. The platform unifies high-performance NVMe flash with cost-effective cloud object storage in a single namespace, automating data tiering to eliminate over-provisioning.