Gnn Disassembly Worldmodel

Real robot disassembly episodes include per-frame constraint graphs, SAM2 segmentation masks, 256D feature embeddings, and full 3D depth information. The dataset was created by Chang Liu of Texas A&M University for a CoRL 2026 project on GNN world models for constraint-aware video generation. It was last updated on Hugging Face on April 11, 2026.

Use Cases

• Training GNN-based world models for robotic disassembly based on per-frame constraint graphs.
• Developing video generation models for robot actions based on synchronized robot states and visual data.
• Researching object segmentation and feature extraction in manipulation tasks based on SAM2 masks and 256D embeddings.
• Analyzing 3D spatial relationships in robotic tasks based on provided point clouds and depth information.

Strengths

• Includes multiple synchronized data modalities: constraint graphs, segmentation masks, feature embeddings, and 3D point clouds.
• Data originates from real robot disassembly episodes using a UR5e and Robotiq 2F-85 hardware.
• The robot is explicitly labeled as a separate agent node with its own mask, embedding, and depth bundle.

Limitations

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

Provenance

Source

Chang Liu (Texas A&M University), hosted on Hugging Face.

Collection Method

Collected from real robot disassembly episodes.

Freshness

Last updated 2026-04-11 23:25:04; freshness should be verified.