Skip to content

Grounding/Grasping Issues on OCID-VLG #3

Description

@sungsuk1m

We evaluated GraspMAS on the OCID-VLG dataset and observed performance issues regarding object grounding and grasp generation.

1. Grounding Failure (GroundingDINO)

The GroundingDINO model frequently fails to ground the target object when allowing up to 5 rounds.

  • Red Rectangle: Model Prediction
  • Green Rectangle: Ground Truth
Image

2. Overly Large Grasp Rectangles

The generated grasp rectangles are often significantly larger than the target object and do not align with the object’s geometry.

  • Red Rectangle: Model Prediction
  • Green Rectangle: Ground Truth
Image

Overall Performance

When running the full pipeline on OCID-VLG, we observed an overall success rate of ~17%, which is lower than expected.

Reproducibility / Verification Request

We have compiled five specific failure cases including their respective prompts. We would appreciate it if the maintainers (or other users) could run these cases and confirm whether the same grounding and grasping behavior is observed.

We would appreciate any insight into whether these results are expected under the current implementation or if additional configuration is required.

Thank you for your assistance.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions