After completing the application deployment of virtual simulation data, Tuosida has once again supplemented the key capability of real machine data collection. This real machine multimodal data acquisition system is a core component of Tuosida's "scene+robot+data+AI model", bridging the data gap between hardware terminals and intelligent algorithms, and building a complete business chain of on-site operations, equipment operation, data precipitation, and intelligent iteration.
The real machine multimodal data acquisition system is based on the self-developed high-precision multimodal data acquisition gripper by Tuosida, and builds an integrated data acquisition platform from real machine acquisition to model iteration. Our self-developed digital gripper has strong versatility and can adapt to various models without the need for robot joint mapping. With lightweight modifications, it can be put into production line use. The device integrates visual inertia and high-frequency sensing solutions to achieve six degrees of freedom high-precision pose tracking and conventional working condition positioningAverage error ≤ 0.25%Synchronize the collection of multidimensional information such as visual images, motion trajectories, end effector postures, and gripper movements, compatible with teaching complex single handed and dual handed processes, to collect and precipitate frontline production data and feed back model iterations, allowing robots to truly "understand the process and know how to work".
The integrated data acquisition platform covers five stages of embodied intelligence data closed-loop, including manual teaching data acquisition, data processing, model training, robot deployment, and feedback iteration. It effectively solves industry pain points such as insufficient adaptability, limited collection accuracy, and weak scene generalization ability of traditional robots, and lays a solid data foundation for the large-scale implementation of industrial embodied intelligence.
Artificial teaching and multimodal acquisition
Settling high-quality data samples
In the first stage, the operator completes industrial tasks such as grasping, handling, and assembly through the data acquisition gripper, and the system synchronously records all dimensional operation and task information, fully retaining the logic and details of manual operation, generating high-quality operation samples that fit the industrial site, and enabling the robot to "understand the process".
Data Processing and Standardization
Improve data processing efficiency and management level
In the second stage, after data collection is completed, the system organizes, aligns, and standardizes multi-source data such as wrist images, motion poses, and gripper states, and encapsulates them into a training dataset. Simultaneously provide visual review tools that support manual viewing, commenting, and usability tagging, providing standardized data input for model training and subsequent iterations. And we will gradually improve our automated data cleaning, quality inspection, classification and archiving, and version management capabilities, further enhancing data processing efficiency and management level.
Model Training and Capability Assessment
Enhance the generalization ability of robots
In the third stage, robot operation strategy training is carried out based on standardized datasets, enabling the model to learn the mapping relationship from visual observation and task instructions to end effector and gripper movements, and continuously optimize the model's capabilities through training validation, effect evaluation, and data supplementation.
Real machine deployment of robots
Implement intelligent homework
In the fourth stage, the high-precision strategy model trained can quickly integrate into various robot control systems, relying on real-time visual perception, robot operating status, and job task instructions to autonomously output precise execution actions, and stably achieve unmanned autonomous execution of typical industrial tasks such as grasping, handling, and assembly, allowing robots to "work" in real factories.
Real machine evaluation and data reflow
Upgrade the feedback model
In the fifth stage, the platform records the entire process and effectiveness of the robot's real machine operation, accurately attributes abnormal and failed scenarios, flows high-quality incremental data back to the collection and training stages, continuously iterates and optimizes the model's operation accuracy and adaptability, and forms a self optimizing data loop.
At this point, Tuosida has established a full process data loop of "manual teaching and collection - data processing and standardization - model training - real machine deployment - feedback optimization", efficiently constructing high-quality industrial datasets, continuously expanding the boundaries of embodied intelligent applications, making embodied intelligent robots more suitable for production processes and better at real-life operations, and helping the intelligent and unmanned transformation and upgrading of the manufacturing industry.
Writing | Xu Jingyu
Review | Tusda Body Intelligent Product Department
