ACME: A Multi-Cultural, Multi-Embodiment Social-Navigation Dataset
Abstract Overview
This paper introduces ACME, a multi-modal social-navigation dataset designed to capture how robot and pedestrian behavior varies across cultural contexts, environments, and robot forms. The dataset was collected by 8 teams across 5 countries using 7 robot embodiments, and includes 29.35 hours of onboard robot data and 43.5 hours of overhead bird's-eye-view pedestrian tracking data. ACME combines egocentric RGB, LiDAR, odometry, calibration data, scenario annotations, and human-verified pedestrian trajectories transformed into metric coordinates. A subset of the dataset also records robot speech during navigation, allowing study of explicit robot-crowd communication in goal-directed social navigation scenarios.
Novelty
What appears most distinctive is the combination of cross-cultural coverage, multiple robot embodiments, and explicit robot speech interactions within a single social-navigation dataset. The work also pairs onboard robot sensing with large-scale, human-verified BEV pedestrian trajectories in metric space, which is uncommon among prior datasets.
Results
The authors show that ACME contains more socially constrained and challenging navigation scenes than recent social-navigation datasets, with a geometric planner exhibiting substantially higher failure rates in ACME despite similar Hausdorff-distance distributions. Benchmarks further show that state-of-the-art vision navigation and pedestrian trajectory prediction models perform worse on ACME than on more established benchmarks, indicating that the dataset captures harder and more diverse behavior. The analyses also reveal context- and location-dependent differences in walking speed, passing side, and personal space.
Key Points
- ACME spans 8 sites in 5 countries and 7 robot embodiments, and provides 29.35 hours of onboard robot data plus 43.5 hours of BEV pedestrian tracking data.
- The dataset includes synchronized multi-modal robot data, human-verified pedestrian trajectories in metric coordinates, scenario and semantic tags, and a subset with robot speech annotations.
- Empirical analyses and benchmarks indicate that ACME is a challenging testbed for social navigation and trajectory prediction, exposing higher planner failure rates and lower model performance than prior benchmarks.