Dataset · ember-lab-berkeley
Large-scale robot pretraining demonstration corpus
ember-lab-berkeley/robocasa365-pretrain-mg
1,615 hours of synthetic video and tabular data across 60 robotics tasks in LeRobot format.
What's inside
Part of the RoboCasa365 collection in standard LeRobot v3.0 format, containing 536,030 episodes (over 116 million frames at 20 fps, totaling 1,615 hours) from three 256x256 h264 cameras. Data generated via MimicGen by augmenting human demonstrations for 60 base tasks (about 10k demos per task) with 720 text instruction variants. Includes video and tabular data.
What you can do with it
- Pretraining a multimodal policy model or VLA (Vision-Language-Action) for robot control
- Training visuomotor skills to execute 60 atomic manipulation tasks via text commands
- Testing imitation learning algorithms on large-scale synthetic data
- Tasks
- Robotics
- Size
- 100M–1B rows
- Downloads this month
- 20.7K
- Likes
- 2
How to load it
from datasets import load_dataset
ds = load_dataset("ember-lab-berkeley/robocasa365-pretrain-mg")