Chain-of-Detection: Enhancing Cross-Granularity Robotic Perception for Object Manipulation.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Cross-granularity object detection is crucial for robotic perception, enabling target identification at various detail levels.
• Traditional methods struggle with the coarse-to-fine detection gap, hindering part association (e.g., cup and handle).
• Vision-language models (VLMs) face fine-grained detection challenges due to limited annotated datasets.

Purpose of the Study:

• To develop a framework for step-by-step detection from coarse recognition to fine-grained localization.
• To address limitations in fine-grained component recognition within existing detectors.
• To automate the generation of fine-grained datasets for improved detector performance.

Main Methods:

• Proposed the chain-of-detection (CoD) framework for guided, step-by-step detection.
• Integrated CoD with Monte Carlo tree search (MCTS) for automated fine-grained dataset generation.
• Eliminated manual labeling requirements through MCTS-driven data synthesis.

Main Results:

• Achieved an average 17.31% improvement in robotic manipulation success rates for common objects.
• Demonstrated a 51.39% improvement for larger object operations.
• Reported approximately 50% improvement in simulated environments.

Conclusions:

• The CoD framework effectively advances cross-granularity detection capabilities.
• Automated dataset generation via MCTS significantly enhances fine-grained detection performance.
• The approach leads to substantial improvements in precise robotic manipulation tasks.