RoboNet
UC Berkeley (BAIR), Stanford, UPenn GRASP, Google Brain · 2019 · datasets.bot · datasets.bot page
One-liner. Open multi-robot manipulation dataset: 15M+ frames of tabletop robot-object interaction from 7 robot platforms and 113 camera viewpoints.
Setup
- Datasets / benchmarks: RoboNet is a large-scale, open dataset for multi-robot learning, containing over 15 million video frames of robot-object interaction collected autonomously across four research labs (UC Berkeley BAIR, Stanford AI Lab, UPenn GRASP, Google Brain Robotics). It aggregates data from 7 robot platforms (Sawyer, Franka Panda, Baxter, Fetch, Google R3, Kuka LBR iiwa, WidowX) recorded from 113 unique camera viewpoints in tabletop settings, with each datapoint storing the camera RGB image, arm pose, force-sensor readings, gripper state and actions. The TFDS release exposes 162,417 trajectories with 5-dimensional action and state vectors. License: CC-BY-4.0. Download: https://www.tensorflow.org/datasets/catalog/robonet.
- Hardware / simulator: Embodiment: fetch, franka_panda, kuka_iiwa, multi, sawyer, widowx. Environment: lab, tabletop. Realness: physical.
Schema
episodes (trajectories) -> steps -> {video (RGB frames), actions (5-dim float32: 3 position + 2 rotation + gripper deltas), states (5-dim float32 EE control state), filename}
Links