Robocasa365 Pretrain Mg: 536,000 Robot Manipulation Demos for 60 Tasks

536,030 episodes of robot manipulation demonstrations generated via the MimicGen scripted augmentation method, totaling 1,615 hours of video. This dataset, part of the RoboCasa365 collection, contains rollouts across 60 atomic tasks, with each task represented by approximately 10,000 demos. It was created by ember-lab-berkeley and last updated in May 2026.

Use Cases

• Pretraining robot manipulation policies based on the 536,000 script-augmented demonstration episodes.
• Benchmarking imitation learning algorithms on the 60 underlying atomic task classes.
• Training vision-language-action models using the natural-language task phrasings paired with multi-camera video data.
• Studying the effects of large-scale synthetic data augmentation via the MimicGen method on robot learning.

Strengths

• Large scale with 536,030 episodes and 116,246,439 video frames.
• High task diversity, covering 60 distinct atomic manipulation tasks.
• Procedurally generated consistency, with approximately 10,000 demonstrations per task created via the MimicGen scripted augmentation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect the biases inherent to the specific human demonstrations and scripted augmentation methods used.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

ember-lab-berkeley via Hugging Face.

Collection Method

Generated by scripted augmentation (MimicGen) from human demonstrations.

Freshness

Last updated 2026-05-11 02:58:06; freshness should be verified.