The industrial AI capability base has landed, and Topstar has completed the key puzzle of embodied intelligent data closed-loop

The scale of industrial embodied intelligence has achieved a key breakthrough. Based on nearly 20 years of manufacturing scene accumulation, Tuosida and Wuwen Zhike have jointly built a data ecosystem, established a complete industrial data iteration loop, and achieved rapid iteration, standardized reuse, and batch landing of embodied intelligent robots.

Wuwen Zhike is the first technology company in the industry to launch a physical AI data base platform. Business covers AI big modelsIndustrial robotIn the four major fields of embodied intelligence, it can output a complete set of data and simulation services, and has won tens of millions of simulation testing orders from multiple leading enterprises.

　　Topstar collaborates with Wuwen Zhike to create a closed-loop platform for physical AI data infrastructure in industrial handling scenariosCovering the entire process from virtual simulation, data annotation, data organization, model training to real machine evaluation, providing core support for the continuous evolution of Tusda's embodied intelligent products.

　　Virtual Simulation Layer: Extreme Compression Debugging Cycle

Based on a self-developed physics engine, a high fidelity digital twin environment is constructed. Multiple training scenarios are generated in batches using Scan2Sim and Gen2Sim, and long tail edge cases are directionally synthesized. The on-site debugging cycle is compressed from "monthly level" to "weekly level".

Data annotation layer: Ten times more efficient, improving quality and efficiency

The simulation data comes with perfect truth values, and the real data adopts automatic pre labeling+manual review mode, which increases efficiency by 10 times.

Data organization layer: Building a sustainable positive data flywheel

Relying on the "data highway" to uniformly access multi-source data from simulation, real machines, and production lines, through automated cleaning, hierarchical classification, and version management, while automatically backflowing simulation amplification of real machine failure cases, a positive data flywheel of "discovering problems supplementary data model upgrade" is formed.

Model Training Layer: Achieving Continuous Autonomous Evolution of Robots

Build a three-level training framework of "perception planning control", achieve rapid implementation through simulation pre training and real machine fine-tuning, and support incremental learning and hot updates, allowing the robot's capabilities to continuously evolve with operation.

Real machine evaluation layer: Comprehensive quantitative verification model capability

Design a embodied simulation evaluation framework, combined with an indicator system and evaluation mechanism covering core dimensions such as task success rate, execution efficiency, and safety in industrial scenarios, to achieve comprehensive and high fidelity quantitative evaluation of model capabilities.

　　

Based on the virtual and real dual drive data base jointly built this time, Tuosida has completed the entire process of data acquisition, simulation expansion, model training, and real machine deployment.All the embodied intelligent robots in the Tuosida series can rely on the platform to complete standardized iterations, breaking free from the limitations of customized research and development in single scenarios, and achieving rapid reuse of technical solutions and cross industry batch replicationOpen up long-term growth space for the large-scale commercial use of robots in multiple scenarios.

　Writing | Xu Jingyu