Robots Watching Robots: Human Pose Estimation Applied to Humanoid Robots
PUBLICATION DATE: 12 December, 2025 PUBLICATION AUTHOR/S: Young, James & Roberts, JonathanThis paper addresses the challenge of adapting human pose estimation (HPE) frameworks to humanoid robots. While HPE has been extensively studied, its direct application to humanoids has proven unreliable due to differences in morphology, symmetry, and appearance. This study makes two contributions. First, it demonstrates the limitations of applying OpenPose to humanoid robots, confirming poor transferability without adaptation. Second, it introduces the first baseline humanoid pose estimation (HuPE) model for the Unitree G1 robot, trained on a custom dataset of 8,410 images, of which 500 were annotated with a 21-keypoint skeleton. The model, developed using YOLOv8, achieves high validation scores on a pilot dataset, including 99.7% precision and 99.5% mAP@50. Qualitative analysis confirms robustness for central landmarks, though appendages and rear-facing views remained challenging. This work establishes a proof-of-concept dataset and model for Unitree G1 HuPE, providing a foundation for expanded datasets, rigorous benchmarking, and practical applications in safety monitoring, collision avoidance, and cooperative manipulation.
RELATED PROGRAM/S:Biomimic Cobots Publication link
View all publications