ACME Dataset: 7 Robots, 5 Countries, 29 Hours of Social Navigation Data

Social robot navigation – getting robots to move safely and politely around people – is notoriously difficult because acceptable behavior varies by culture, environment, and robot form. Existing datasets are mostly collected in single locations with one robot type, missing the diversity that real-world deployment demands. To bridge this gap, a team of over 40 researchers from Carnegie Mellon, IBM Research, George Mason, University of Bonn, National University of Singapore, and others collaborated on ACME: A Cross-cultural, Multi-Embodiment dataset. The dataset spans 8 sites across Japan, USA, Singapore, India, and Spain, using 7 different robot embodiments including wheeled bases, quadrupeds, and humanoids. In total, ACME provides 29.35 hours of onboard sensor data (LiDAR, cameras, IMU) and 43.5 hours of overhead pedestrian tracking video. Crucially, the robots were programmed to proactively interact with crowds using speech, creating realistic goal-driven social scenarios.

ACME stands out by explicitly labeling pedestrian trajectories and interactions, enabling researchers to train models that predict human movement and adapt navigation policies in real time. The dataset also includes 3D and 2D scene features, odometry, and detailed human annotations. The authors' analysis shows that ACME captures a broader distribution of pedestrian behaviors than prior datasets like THÖR or Social-Nav. With its multi-cultural and multi-embodiment design, ACME aims to be a standard benchmark for social navigation research. The paper was submitted to IJRR in June 2026 and is available on arXiv with code and data links. For engineers building service robots, warehouse assistants, or autonomous delivery vehicles, ACME offers a richer, more realistic training resource that should lead to safer and more socially acceptable robot behavior across different regions.