Humanoid locomotion is mostly solved. But whole-body manipulation in messy, real-world homes? Still one of the biggest bottlenecks in robotics. Laboratory data doesn't cut it. To build "robot butlers," models need to train on the chaotic edge cases of the real world. That’s where HIW-500 (Humanoids-in-the-Wild) comes in—the largest open-source humanoid teleop dataset collected entirely in real homes. Developed by @BitRobotNetwork alongside @LeRobotHF @huggingface and Unitree Robotics, HIW-500 offers the massive data scale required to solve real-world generalization: • 500+ hours of whole-body data • 23,000+ individual episodes • 10+ TB of raw data across 12 authentic homes • 10+ long-horizon tasks (sweeping, tidying, object manipulation) The data was gathered via a fleet of Unitree G1 humanoids utilizing a rigorous sensor stack: • Stereo head cameras (480p, 30 FPS) • Dual infrared (IR) stereo wrist cameras to eliminate visual occlusion during manipulation • Full 29-DoF kinematic states, IMU, and odometry Infrastructure hurdle solved: Hugging Face's LeRobot team re-encoded the raw data, compressing it from ~10TB down to ~2TB with ZERO loss of fidelity. Because the Unitree G1 is natively supported in LeRobot, researchers can pull these real-home trajectories and start training imitation or VLA policies today. A massive open-source step toward hitting the industry’s "80/80" benchmark (achieving an 80% task completion rate across 80% of unfamiliar, real-world scenes) for general-purpose robots.