Bi3 dataset brings biperson, bicultural data to social robot navigation
74 participants, 10.5 hours, 5 algorithms across two continents...
Researchers from the University of Michigan, LAAS-CNRS (France), and other institutions have released Bi3, a comprehensive dataset for social robot navigation. The dataset stands out by capturing close-range interactions between a robot and two humans in constrained lab environments, using five different navigation algorithms and two robot platforms. Recruited from two sites in the USA and France, 74 participants contributed 10.5 hours of ground-truth motion tracks, RGB video, and subjective impressions of robot performance.
Bi3's design deliberately generates dense, ambiguous social encounters—unlike prior datasets that avoid tight spaces. Analysis shows Bi3 offers uniquely high interaction density and human velocity variability, making it a benchmark for modeling human motion prediction and robot control in crowded settings. Accepted at ICRA 2026, the dataset is publicly available for training socially-aware robot navigation systems.
- 74 participants recruited from USA and France, providing bicultural social dynamics
- 10.5 hours of data with two robot platforms and five navigation algorithms
- Focuses on close-encounter scenarios between two humans and a robot in constrained spaces
Why It Matters
Bi3 enables more realistic, culturally-aware training data for robots navigating dense human crowds.