Bridge-CoT: Robot Manipulation Data with Chain-of-Thought Annotations

Bridge-CoT is a dataset of 35,357 samples for robot manipulation, derived from BridgeDataV2. Each sample pairs a scene image with a task description and includes structured VLM-generated annotations for object detection, spatial relations, and subgoal decomposition. The dataset was created by CliffKai and was last updated on Hugging Face in April 2026.

Use Cases

• Training vision-language models for task understanding based on scene images and task descriptions.
• Developing robotic planning algorithms based on the structured subgoal decomposition annotations.
• Benchmarking object detection and spatial reasoning models using the VLM-generated annotations.
• Studying chain-of-thought reasoning for physical tasks based on the provided annotations.

Strengths

• Contains 35,357 samples, providing a substantial collection for model training.
• Each sample includes a scene image paired with a task description and multiple structured annotation types.
• Annotations are generated by a Vision-Language Model, providing detailed chain-of-thought reasoning.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect bias inherent to its source, BridgeDataV2.

Provenance

Source

Derived from BridgeDataV2.

Collection Method

VLM-generated annotations added to existing robot manipulation data.

Freshness

Last updated 2026-04-02 05:24:41; freshness should be verified.