IA-Bench: Human-Annotated Interacted Objects for Robot Manipulation Subtasks

IA-Bench (Interacted-Object Benchmark) provides human ground-truth annotations of interacted objects for robot manipulation subtasks. Each sample includes a full subtask video clip, gripper proprioception data aligned to frames, a language instruction, and pixel-coordinate bounding boxes for the initial and target object. The dataset is authored by irl-kit and was last updated on June 18, 2026.

Use Cases

• Training object detection models for robotic tasks based on the annotated bounding boxes.
• Developing models for instruction-following in manipulation based on the paired video and language data.
• Benchmarking robotic perception systems using the human-annotated ground-truth object states.
• Analyzing gripper proprioception patterns aligned with visual frames for manipulation subtasks.

Strengths

• Includes human ground-truth annotations, which are a high-quality signal for model training.
• Provides aligned multimodal data: video clips, proprioception, language instructions, and bounding boxes per sample.
• Last updated on 2026-06-18, suggesting recent maintenance.

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

irl-kit

Collection Method

Human ground-truth annotations.

Freshness

Last updated 2026-06-18 12:13:19.