Task-Oriented Dexterous Grasp Data With Language Instructions

DexTOG provides assets for learning dexterous robot grasps conditioned on language descriptions of tasks. The dataset, created by robotflow, was published in April 2025 to accompany the associated research paper. It contains simulation-based data linking grasp poses, object interactions, and natural language instructions.

Use Cases

• Train a model to predict stable grasp poses from language instructions and object geometry features.
• Benchmark reinforcement learning agents for dexterous manipulation using simulation state and reward data.
• Develop vision-language models that map scene observations and text commands to joint torque sequences.
• Analyze the relationship between grasp type labels and task success metrics across different objects.

Strengths

• Data is directly linked to a peer-reviewed research paper (DexTOG) published on arXiv.
• Assets are hosted on a version-controlled platform with a specific last update date of 2025-04-08.

Limitations

• The dataset size, row count, and specific feature dimensions are not documented.
• Data is generated in simulation, which may introduce a sim-to-real gap for real-world robotic application.

Provenance

Source

robotflow organization on Hugging Face, associated with the DexTOG research project.

Collection Method

Likely generated through robotic simulation environments as part of the described research methodology.

Freshness

Last updated on 2025-04-08.