LivUMI Bimanual Robot Manipulation Data with Multiview Video

LivUMI contains imitation learning data for a dual-arm robot, structured in episodes with a LeRobot v2.1-compatible layout. The dataset includes multiview camera streams, depth information, robot state, and temporal indexing. It was authored by lingsheng-ai and last updated in April 2026.

Use Cases

• Train policies to predict dual-arm end-effector poses and gripper values from multiview camera images and depth data.
• Benchmark temporal action segmentation models using frame indices and timestamps across long-horizon manipulation episodes.
• Develop multimodal encoders that fuse non-image state fields (like joint positions) with external video references for task representation.
• Analyze task structure and performance by linking episodes to textual task descriptions via the tasks.jsonl metadata.

Strengths

• Data is organized into episodes compatible with the established LeRobot v2.1 codebase layout.
• Includes multiple synchronized data modalities: camera streams, depth, robot state, and timestamps.

Limitations

• The available description is for a sample version; the full dataset's size, row count, and scope are unknown.
• Specific column names, data volume, and potential class balance for tasks are not provided.

Provenance

Source

lingsheng-ai on Hugging Face.

Collection Method

Collected from a dual-arm robot performing manipulation tasks, likely for imitation learning.

Freshness

Last updated on 2026-04-08.