HABIT (Human Aware Behavior and Interaction Training dataset)
Human-Robot Interaction dataset with 10K+ teleop trajectories (160+h) on bimanual Franka Research 3
| Year | 2026 |
|---|---|
| Episodes | 10,563 |
| Total hours | 164 |
| Embodiments | franka_panda |
| Modalities | ee_pose, language, proprioception, rgb |
| Task categories | manipulation |
| License | CC-BY-4.0 |
| Access | open — commercial use permitted |
| Maintainer | Config |
HABIT (Human-Aware Behavior and Interaction Training dataset) is a large-scale robot manipulation dataset collected in human-present environments, where a co-present human shares the workspace and interacts with the robot in every episode.
- 10,563 episodes, 164 hours of bimanual manipulation across 60 tasks
- Organized into three human-robot interaction roles (from the HRI literature):
Collaborator (jointly do one task), Coworker (separate tasks, shared space), Supervisor (human directs the robot)
- Collected with a reactive interaction protocol (the robot responds to the human's
actual actions, not a memorized routine) and designed to elicit human-aware behaviors: yielding, temporal adaptation, gesture grounding
- 5 synchronized RGB cameras (3 robot-side, 2 human-side) at 640×480, 10 Hz
- Action spaces: joint-space, Cartesian space, and grippers
- Language annotations : provide both high-level and low-level instruction per frame (both for human and robot)