Sign in to view source links and access this dataset
Description
Robocasa365 Pretrain Composite is a dataset of human teleoperation demonstrations for 235 composite, multi-step tasks, with roughly 100 demonstrations per task. It contains 24,687 episodes and 27,610,913 frames, representing the long-horizon half of the RoboCasa365 collection's 482-hour pretraining dataset. The dataset is provided by ember-lab-berkeley and was last updated in May 2026.
Use Cases
Training imitation learning models based on the 24,687 human teleoperation episodes.
Developing or benchmarking robotic policies for long-horizon tasks based on the 235 composite task classes.
Pretraining vision-language-action models based on the multimodal data likely containing camera feeds and natural language task phrasings.
Strengths
Large scale with 24,687 episodes and 27,610,913 frames, equivalent to approximately 383 hours of video at 20 fps.
Focus on long-horizon, multi-step tasks with 235 distinct underlying task classes.
Includes multiple natural language phrasings for tasks, with 4,428 total phrasings noted.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Specific file formats, data size, and license information are unknown from the provided metadata.
Provenance
Source
ember-lab-berkeley
Collection Method
Human teleoperation data collection.
Freshness
Last updated 2026-05-11 02:37:38.
License is unknown; users must verify terms of use before downloading.