A new open-source library called LeRobot has been launched by Hugging Face in collaboration with researchers from the University of Oxford, marking a significant milestone in robotics development. Dubbed the “PyTorch of robotics,” LeRobot is designed to enable end-to-end integration across the entire robotics stack, aiming to accelerate the shift from traditional equation-based control systems to data-driven, generalizable robotic intelligence. LeRobot supports learning from large-scale, multimodal datasets—including text, video, and sensor data—and is compatible with a wide range of robotic hardware. It is built to advance robotics toward more adaptable, general-purpose systems capable of handling diverse tasks across different platforms. One of its core innovations is LeRobotDataset, a standardized, native data format specifically designed for robotics. This format addresses the growing need for consistent, structured, and interoperable data in robot learning. LeRobotDataset stores not only raw data such as images and robot state streams but also rich metadata—like task descriptions provided by human operators, the type of robot used, and technical details such as frame rates during data collection. This unified structure enables seamless handling of multimodal and time-series data, while offering direct integration with PyTorch and the broader Hugging Face ecosystem. The format is also extensible, allowing users to customize it for their own use cases, and it already supports data from a variety of open-source platforms, including the SO-100 robotic arm, ALOHA-2 manipulator, human-like robotic arms, and even simulated environments and autonomous vehicle datasets. The library also includes a collection of state-of-the-art algorithms for robot learning, with efficient implementations in PyTorch, along with built-in support for experiment tracking and reproducibility. LeRobot introduces a customizable inference stack that decouples action planning from action execution, enhancing flexibility and modularity in robot control. According to the accompanying research paper (arXiv:2510.12403), the field of robotics is at a turning point, driven by rapid advances in machine learning and the increasing availability of large-scale, real-world robot data. The paper outlines a clear evolution: from foundational methods like reinforcement learning and behavior cloning, to more advanced, generalizable, language-conditioned vision-language-action models capable of transferring skills across tasks and robotic platforms. Traditional robotics relied heavily on physics-based modeling and hand-coded rules. However, modern approaches increasingly leverage data-driven techniques. While reinforcement learning offers powerful learning-through-interaction capabilities, it suffers from poor sample efficiency, safety risks in real-world deployment, and complex reward engineering. To overcome these challenges, newer hybrid methods such as Human-in-the-Loop Reinforcement Learning (HIL-SERL) incorporate human guidance, prior datasets, and learned reward classifiers to make real-world training more feasible. Meanwhile, imitation learning has gained traction as a safer alternative, enabling robots to learn from limited expert demonstrations without requiring explicit rewards. As these techniques mature, they have paved the way for single-task policies that can be generalized through few-shot or zero-shot learning using language-conditioned models. Recent breakthroughs—such as the π0 model by Physical Intelligence and SmolVLA—have demonstrated remarkable cross-platform generalization by leveraging pre-trained backbones and generative techniques like flow matching. The paper emphasizes that the emergence of large, publicly available datasets and standardized, easy-to-use model architectures has been critical to this progress. LeRobot is positioned as a key enabler of this trend, serving as a central hub that unifies data, models, and tools for the robotics community. By reducing the barriers to entry and minimizing redundant development, LeRobot empowers researchers, engineers, and even hobbyists to train and deploy robot policies with just a few lines of code. It represents a bold step toward a future where robotics becomes more accessible, generalizable, and affordable—transforming the field from closed, proprietary systems into open, collaborative platforms. Reference: https://arxiv.org/pdf/2510.12403 Editorial & Layout: He Chenlong