Physical Intelligence Builds Adaptable Robot Intelligence with Anyscale

In an industry where most companies build robots for specific tasks in controlled environments, Physical Intelligence is taking a different approach: creating truly adaptive robotics models that can function across any hardware platform, task, or environment. Training models that function across diverse robot embodiments – including one-arm or two-arm, mobile or stationary, and dexterous hands or simple grippers – generates terabytes of complex, heterogeneous robotics logs per day. This requires a solution that can handle such large data processing needs, including complex data types like video, and that can be implemented efficiently. Physical Intelligence uses Ray, with its clear and intuitive API for distributed Python computing, on Anyscale, which offers the full power of Ray.

LinkThe Challenge: Distributed Data Processing at Scale

From the earliest days at Physical Intelligence, it was clear that leveraging distributed infrastructure to process the vast quantities of data collected each day was an absolute must. To focus on their differentiated research, the team wanted to fully leverage Ray’s distributed computing engine.

"We want to train a model that can do any task on any robot – and do it well. Our belief is that by training our models on diverse data, they will generalize better and outperform other models trained only on narrower, specialized domains."

Adrian Li-Bell | Member of Technical Staff @ Physical Intelligence

LinkFinding a Solution: Managed Ray on Anyscale

Physical Intelligence chose Anyscale's managed platform. This solution delivered enterprise-grade distributed computing power within an easy-to-use developer environment.

LinkProcessing Multimodal Data at Scale

Physical Intelligence collects terabytes of sensor data daily from robots with varying hardware configurations, tasks, and environments. This quantity – and variety – of heterogeneous data created a critical data processing challenge: transforming massive amounts of heterogeneous sensor data into formats optimized for model training.

The team required a solution that could efficiently coordinate processing across CPU and GPU resources, reducing both time and infrastructure costs. Anyscale delivered:

• Heterogeneous data processing across CPUs and GPUs for optimal performance

• Cloud-based processing on AWS and GCP

• Efficient conversion of complex robotics logs into training-ready tensor formats

LinkEliminating Infrastructure Barriers

The managed Anyscale platform delivered immediate benefits:

• Direct access to Ray experts for infrastructure optimization and troubleshooting

• A managed platform with automatic scaling that eliminated manual resource provisioning

• Tooling for debugging and observability, including log management and monitoring of resource utilization

The improved reliability and operational stability provided by Anyscale enabled the team to redirect valuable engineering resources toward advancing their robotics research.

LinkDeveloper Experience That Accelerates Research

Anyscale provided:

• Intuitive interfaces that abstract away infrastructure complexity

• On-demand scaling for both production pipelines and ad-hoc experiments

• Seamless integration with familiar development environments

"With Anyscale, our researchers can just write code without worrying about the underlying infrastructure. The development-to-deployment workflow is smooth and intuitive, which means we can iterate faster on our models."

Adrian Li-Bell | Member of Technical Staff @ Physical Intelligence

Anyscale's Workspaces, an interactive development environment backed with managed Ray clusters, allows researchers to instantly build and reproduce cloud jobs from VS Code – making distributed computing accessible without specialized Ray expertise.

LinkThe Future of Robotics is Embodiment-Agnostic

With Anyscale, Physical Intelligence has transformed how they process terabytes of robotics sensor data, eliminated infrastructure management burdens, and accelerated research through an intuitive developer experience. This foundation has positioned them to pursue their ambitious vision of truly hardware-agnostic robotics.

For this vision to become reality, Physical Intelligence needs infrastructure that can handle not just large volumes of data, but also provide the low-latency processing that real-time learning requires. Anyscale's flexible scaling and high-performance distributed computing provide the foundation they need to pursue these ambitious goals.

With these tools, Physical Intelligence can focus on building the future of robotics that can adapt to any task, any environment, and any hardware configuration – bringing us closer to truly intelligent, versatile machines that can work alongside humans in the real world.