DiffuseLoco is a novel framework designed to learn multi-skill locomotion control for legged robots from offline datasets. It leverages diffusion models to capture the multi-modality in the dataset, enabling the robot to perform a diverse set of agile locomotion skills, including bipedal walking and quadrupedal movements, within a single policy. The framework addresses challenges such as real-time control, skill transitioning, and robustness to environmental variations. Key contributions include: 1. **State-of-the-Art Multi-Skill Controller**: DiffuseLoco learns a robust controller that can perform various agile locomotion skills, including bipedal walking and quadrupedal movements, in a unified policy. 2. **Scalable Offline Learning**: It directly learns from diverse offline datasets, demonstrating the benefits of offline learning at scale for locomotion skills in real-world scenarios. 3. **Real-World Validation**: Extensive real-world experiments show improved stability and velocity tracking compared to baselines, with free transitions between skills and robustness against environmental changes. The framework consists of three stages: data collection, training, and deployment. Data collection involves gathering demonstrations from various sources, training the policy using a diffusion model, and deploying it on real-world robots. The diffusion model, specifically Denoising Diffusion Probabilistic Models (DDPM), is used to generate sequences of low-level actions conditioned on robot states and goals. The framework also incorporates receding horizon control (RHC) for real-time control and delayed inputs to handle large models efficiently. DiffuseLoco's performance is evaluated through benchmarking and ablation studies, showing superior stability, velocity tracking, and skill transitioning compared to existing methods. The framework's scalability and versatility make it a promising approach for learning-based control in legged locomotion.DiffuseLoco is a novel framework designed to learn multi-skill locomotion control for legged robots from offline datasets. It leverages diffusion models to capture the multi-modality in the dataset, enabling the robot to perform a diverse set of agile locomotion skills, including bipedal walking and quadrupedal movements, within a single policy. The framework addresses challenges such as real-time control, skill transitioning, and robustness to environmental variations. Key contributions include: 1. **State-of-the-Art Multi-Skill Controller**: DiffuseLoco learns a robust controller that can perform various agile locomotion skills, including bipedal walking and quadrupedal movements, in a unified policy. 2. **Scalable Offline Learning**: It directly learns from diverse offline datasets, demonstrating the benefits of offline learning at scale for locomotion skills in real-world scenarios. 3. **Real-World Validation**: Extensive real-world experiments show improved stability and velocity tracking compared to baselines, with free transitions between skills and robustness against environmental changes. The framework consists of three stages: data collection, training, and deployment. Data collection involves gathering demonstrations from various sources, training the policy using a diffusion model, and deploying it on real-world robots. The diffusion model, specifically Denoising Diffusion Probabilistic Models (DDPM), is used to generate sequences of low-level actions conditioned on robot states and goals. The framework also incorporates receding horizon control (RHC) for real-time control and delayed inputs to handle large models efficiently. DiffuseLoco's performance is evaluated through benchmarking and ablation studies, showing superior stability, velocity tracking, and skill transitioning compared to existing methods. The framework's scalability and versatility make it a promising approach for learning-based control in legged locomotion.