July 2026 Summaries

The Anyscale Physical AI Skill is a new tool designed to streamline the development of robotics and autonomous-driving systems by leveraging Ray and Anyscale. It provides a comprehensive solution for building physical AI workloads, encompassing everything from scoping and decision-making to generating necessary code, dependencies, and configurations. The skill supports various workload classes, including offline and online reinforcement learning (RL), and integrates with models and environments such as Alpamayo, Cosmos, and AlpaSim. It aims to overcome the operational complexities of physical AI workloads, such as the need for heterogeneous computing resources and the inherent challenges of managing dependencies and simulators. The Physical AI Skill ensures that workloads are built from validated, efficient patterns to avoid common pitfalls and to enhance scalability and reliability. It generates open-source code, allowing users full control and flexibility while maintaining compatibility with existing systems. By turning a concept into a requirements-driven implementation, the skill facilitates the creation of a data flywheel for iterative policy training, simulation, and deployment, ultimately advancing the capabilities of robotics and autonomous systems on the Anyscale platform.

Modern AI applications, which are increasingly resource-intensive, often face challenges such as workload instability and job failures due to resource contention. To address these issues, Ray's Resource Isolation feature was introduced, leveraging Linux kernel control groups (cgroup-v2) to enhance stability by isolating critical system processes from workload processes. This isolation ensures that workloads remain stable and predictable, even under resource contention, by implementing constraints like memory.low and memory.high, which optimize resource allocation and prevent undesired kernel Out-Of-Memory (OOM) kills. Additionally, the feature includes a workload-aware memory monitor that prioritizes preserving task progress and improves system stability, significantly reducing node and job failures. By solving both memory and CPU contention issues, Ray with Resource Isolation delivers improved performance and reliability for tasks such as large-scale video data processing pipelines, showcasing significant completion time improvements and eliminating kernel OOM kills. This advancement underscores Ray's commitment to providing robust solutions for managing the growing demands of modern AI workloads.