Edge-intelligent vision-based robotic manipulation for real-time pick-and-place in dynamic industrial environments

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Vision-based robotic manipulation in dynamic industrial settings requires robust perception, rapid decision-making, and reliable closed-loop control under uncertainty.
• Existing systems often face challenges with latency and robustness in dynamic environments, particularly with occlusions, lighting changes, and object pose uncertainty.

Purpose of the Study:

• To present an edge-intelligent robotic manipulation framework for pick-and-place tasks.
• To improve robustness, reduce latency, and enhance generalization capabilities in vision-based robotic manipulation.
• To integrate perception, calibration, motion planning, and control into a unified closed-loop architecture.

Main Methods:

• Developed a simulation-based closed-loop architecture combining perception, calibration, motion planning, and control.
• Utilized a lightweight Convolutional Neural Network (CNN)-based perception module and calibration-aware SE(3) transformation refinement.
• Employed a data-centric enrichment technique for improved generalization and edge-enabled execution logic for reduced latency.
• Assessed the framework using Monte Carlo simulations and trials in diverse industrial contexts.

Main Results:

• The proposed framework demonstrated improved task success rates, reduced latency, and enhanced robustness compared to existing robotic manipulation baselines.
• The data-centric enrichment technique improved generalization under conditions of occlusion, light change, and object pose uncertainty.
• Consistent performance was observed between high-fidelity simulation and a small-scale real-world validation (18 grip trials).

Conclusions:

• The integration of edge intelligence with closed-loop robotic control proves effective for enhancing robotic manipulation.
• The framework offers a promising approach for robust and efficient vision-based robotic manipulation in dynamic industrial settings.
• The study validates the framework's efficacy through simulation and limited physical testing, demonstrating its transferability.