Robocasa365 Target Atomic: Human Teleoperation Data for 18 Kitchen Tasks

9,126 episodes of human teleoperation data recorded in 10 held-out target kitchens. The dataset contains 2,231,347 frames of video and action data for 18 underlying task classes, represented by 231 natural-language phrasings. It was created by ember-lab-berkeley and last updated on May 11, 2026.

Use Cases

• Train imitation learning models based on human teleoperation demonstrations.
• Benchmark robotic manipulation performance on 18 atomic kitchen tasks.
• Study generalization across 10 held-out target kitchen environments.
• Develop multi-task or few-shot learning models based on the 231 natural-language task phrasings.

Strengths

• Large-scale collection with 9,126 episodes and over 2.2 million frames.
• Focused on 18 atomic tasks with 500 demonstrations per task.
• Includes data from 10 distinct, held-out target kitchens for testing generalization.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last updated 2026-05-11 02:30:28; freshness should be verified.

Provenance

Source

ember-lab-berkeley

Collection Method

Human teleoperation recordings.

Freshness

Last updated May 11, 2026.