LA-33K: Language-Action Pretraining Data for Robot Manipulation

33,116 short, atomic manipulation episodes collected for pretraining vision-language-action models. The dataset contains 1.52 million frames and 9,532 language tasks, formatted for the LeRobot v2.1 framework and derived from Franka robot demonstrations. It was created by MINT-SJTU and last updated on July 8, 2026.

Use Cases

• Pretrain vision-language-action models based on the 33,116 short, atomic manipulation episodes.
• Develop robot instruction-following capabilities based on the 9,532 language tasks.
• Benchmark robotic manipulation algorithms using the 1.52 million frames of episodic data.
• Train models for multimodal robotic control using the LeRobot v2.1 format.

Strengths

• Contains 33,116 manipulation episodes, providing a substantial collection for pretraining.
• Includes 1.52 million frames and 9,532 distinct language tasks, offering multimodal content.
• Formatted for the LeRobot v2.1 framework, which suggests a specific application standard.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Last updated 2026-07-08 16:36:03; freshness should be verified.
• The dataset's specific source and collection methodology are not detailed in the provided description.

Provenance

Source

MINT-SJTU

Collection Method

Derived from Franka robot demonstrations, likely containing recorded episodes.

Freshness

Last updated July 8, 2026.