DiffuseLoco is a scalable framework that leverages diffusion models to learn legged locomotion control from offline datasets. It enables real-time control of diverse locomotion skills on robots in the real world. The framework learns a state-of-the-art policy that can perform a diverse set of agile locomotion skills with a single policy, exhibiting robustness in the real world and versatility to various sources of offline data. DiffuseLoco addresses challenges in learning from offline data, including generating diverse skills with the same goals. It is capable of reproducing multimodality in performing various locomotion skills, zero-shot transfer to real quadrupedal robots, and deployment on edge computing devices. DiffuseLoco demonstrates free transitions between skills and robustness against environmental variations. Through extensive benchmarking in real-world experiments, DiffuseLoco exhibits better stability and velocity tracking performance compared to prior reinforcement learning and non-diffusion-based behavior cloning baselines. The design choices are validated via comprehensive ablation studies. This work opens new possibilities for scaling up learning-based legged locomotion controllers through the scaling of large, expressive models and diverse offline datasets. DiffuseLoco is a novel approach that emphasizes learning agile legged locomotion skills at scale by solving two aforementioned challenges: offline learning from various data sources and the ability to learn a set of diverse skills. The framework leverages expressive diffusion models to effectively learn the multi-modalities that exist in the diverse offline dataset without manual skill labeling. Once trained, our controllers can execute robust locomotion skills on real-world legged robots for real-time control. The primary contributions of this work include: 1) A state-of-the-art multi-skill controller, leveraging expressive diffusion models, that learns agile bipedal walking and various quadrupedal locomotion skills within a single policy and can be deployed zero-shot on real-world quadrupedal robots. 2) A novel framework that directly learns from a diverse offline dataset for real-time control of legged robots, showing the benefits and potentials of offline learning at scale for locomotion skills practically in a real-world scenario. 3) Extensive real-world validation showing higher stability and lower velocity tracking errors compared to baselines, while demonstrating multi-modal behaviors with skill transitioning and robustness on terrains with varying ground frictions. This work opens up the possibility of leveraging large-scale learning to create diverse and agile multi-skill controllers for legged locomotion from offline datasets. For the first time, we show that it is feasible to zero-shot transfer such a diverse locomotion policy learned from a static dataset to real-world applications. This approach offers a scalable and versatile framework for learning-based control, allowing for continuous expansion of the dataset and integration of diverse skills from various data sources. Codebase and checkpoints will be open-sourced upon the acceptance of this work.DiffuseLoco is a scalable framework that leverages diffusion models to learn legged locomotion control from offline datasets. It enables real-time control of diverse locomotion skills on robots in the real world. The framework learns a state-of-the-art policy that can perform a diverse set of agile locomotion skills with a single policy, exhibiting robustness in the real world and versatility to various sources of offline data. DiffuseLoco addresses challenges in learning from offline data, including generating diverse skills with the same goals. It is capable of reproducing multimodality in performing various locomotion skills, zero-shot transfer to real quadrupedal robots, and deployment on edge computing devices. DiffuseLoco demonstrates free transitions between skills and robustness against environmental variations. Through extensive benchmarking in real-world experiments, DiffuseLoco exhibits better stability and velocity tracking performance compared to prior reinforcement learning and non-diffusion-based behavior cloning baselines. The design choices are validated via comprehensive ablation studies. This work opens new possibilities for scaling up learning-based legged locomotion controllers through the scaling of large, expressive models and diverse offline datasets. DiffuseLoco is a novel approach that emphasizes learning agile legged locomotion skills at scale by solving two aforementioned challenges: offline learning from various data sources and the ability to learn a set of diverse skills. The framework leverages expressive diffusion models to effectively learn the multi-modalities that exist in the diverse offline dataset without manual skill labeling. Once trained, our controllers can execute robust locomotion skills on real-world legged robots for real-time control. The primary contributions of this work include: 1) A state-of-the-art multi-skill controller, leveraging expressive diffusion models, that learns agile bipedal walking and various quadrupedal locomotion skills within a single policy and can be deployed zero-shot on real-world quadrupedal robots. 2) A novel framework that directly learns from a diverse offline dataset for real-time control of legged robots, showing the benefits and potentials of offline learning at scale for locomotion skills practically in a real-world scenario. 3) Extensive real-world validation showing higher stability and lower velocity tracking errors compared to baselines, while demonstrating multi-modal behaviors with skill transitioning and robustness on terrains with varying ground frictions. This work opens up the possibility of leveraging large-scale learning to create diverse and agile multi-skill controllers for legged locomotion from offline datasets. For the first time, we show that it is feasible to zero-shot transfer such a diverse locomotion policy learned from a static dataset to real-world applications. This approach offers a scalable and versatile framework for learning-based control, allowing for continuous expansion of the dataset and integration of diverse skills from various data sources. Codebase and checkpoints will be open-sourced upon the acceptance of this work.