From Data Benchmark to Safe Execution: The Closed-Loop Ecosystem and Evolution of Embodied Intelligence

I was honored to give a talk at the “紫器东来” Academic Workshop held by the School of Instrument Science and Engineering, Southeast University.

This talk delves into the cutting-edge vision of “Closed-Loop Ecosystems for Embodied Intelligence,” systematically outlining how to drive robots from virtual simulation to agile, safe operations in the real world. It focus on three core challenges: generalizing perception from multi-task embodied data, robust decision-making for long-range complex tasks, and high-security physical execution in complex dynamic environments. Specific content includes: developing a data benchmark and Sim2Real transfer for industrial embodied intelligence RoCoChallenge tasks targeting sub-millimeter level dual-arm manipulation; constructing the general zero-shot robot operation framework RoboDexVLM, enabling long-range task decomposition and precise action generation for robotic arm dexterity using multimodal large models; proposing the Unified Decision and Motion Control framework (UDMC) series of works to address the challenges of generalization and robustness in mobile navigation; creating an original, integrated closed-loop ecosystem for embodied intelligence that fuses “optimization, learning, and large models,” grounded in data benchmarks and secured by safe execution, driving cognitive upgrades; and finally, introducing the latest real-world applications of this technology in industrial collaborative assembly and high-security autonomous systems.

For more information, please refer to the SEU Workshop Page.