EgoSPT: Egocentric Manipulation Trajectories for Robot Learning

EgoSPT is a dataset of 11,515 episodes of robot manipulation trajectories collected from three scenes. It contains egocentric RGB videos aligned with end-effector poses and gripper widths for vision-conditioned trajectory prediction. The dataset was created by JackYFL233 and is hosted on Hugging Face.

Use Cases

• Training vision-based robot control models based on egocentric RGB video.
• Predicting future end-effector trajectories based on current visual and pose data.
• Developing multimodal models that fuse video and action data for robotic tasks.
• Benchmarking algorithms for manipulation trajectory prediction in defined scenes.

Strengths

• Contains 11,515 processed episodes, providing a substantial number of training examples.
• Includes multiple aligned data modalities: RGB video, end-effector poses, gripper widths, and valid-frame masks.
• Covers 112 distinct task folders across three different scenes.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

JackYFL233 on Hugging Face.

Collection Method

Collected for vision-conditioned trajectory prediction, likely from robotic manipulation experiments.

Freshness

Last updated 2026-05-14 17:30:58; freshness should be verified.