Embodied data collection moves from collection plants to everyday life
Currently, the data needed for embodied intelligence remains limited to laboratories and centralized data collection scenarios. It is difficult to systematically cover the diverse operations and physical interactions in the real world. Looking back at the development of autonomous driving, we see that as systems moved from the laboratory to real roads, data collection moved from small-scale, controlled approaches to large-scale accumulation in real-world scenarios, which facilitated the rapid evolution of model capabilities.
Based on this experience, Qiongche Intelligence was officially launched RoboPocket – Through smartphones and apps, every user becomes a participant in data collection, completing tasks and uploading data, enabling lightweight, controllable, and high-quality data collection.. This extends data collection to a wider range of real-world environments and allows more users to participate in task collection. Continuously generate high-quality, usable data that is lightweight and controllable, building a more realistic and reliable data foundation for embodied intelligence models.
RoboPocket – Enabling everyone to collect high-quality data
Robopocket is a high-quality, user-friendly data collection solution that can be activated right out of your pocket.
Portable and easy to use: Highly integrated and high quality data collection device
Data collection no longer relies on complex specialized equipment. With your phone you can unlock unlimited high quality collections. Robopocket is lightweight and ready to use right out of the box. Users can connect their devices and start collecting with just a light touch on their phone.
Caption: Unpacking and quick start
This lightweight data collection solution doesn't sacrifice accuracy.
Mobile phones such as the iPhone are equipped with both an RGB camera and a depth camera (LiDAR). Compared to pure visual SLAM, it is a multi-sensor fusion (vision, depth, IMU) solution. With this approach, the collection accuracy is higher than pure VSLAM. Compared to infrared positioning, there is no need to drag base stations while maintaining a high degree of integration. The removable and replaceable fisheye lens provides an ultra-wide field of view, while the native lens has excellent image quality and can withstand challenges in real-world scenarios such as strong and weak light. It maintains good performance even in dimly lit rooms and outdoor environments exposed to direct sunlight.
Caption: Fisheye lens, super wide field of view
RoboPocket uses the mature capabilities of mobile phones to complete SLAM positioning and mapping. Compared to the original UMI, which required manual calibration and repeated operations, one of RoboPocket's core values is to “make mapping a seamless experience for users.”
Caption: Rapid mapping
Robopocket also acts as an intelligent hub to ensure collection quality in real time
In the era of large-scale reification models, data quality is as important as data quantity.
Traditional UMI-style data collection solutions have always faced issues such as: The impossible triangle of data quality, portability and ease of use, and post-processing pressures. If you want to ensure data quality, you usually have to sacrifice convenience. To improve the efficiency of quality feedback at the data collection site, collection devices are connected to computers, but this method is destined not to reach thousands of households. Alternatively, a small device that integrates storage and data collection is required, which imposes the burden of data processing after collection, resulting in low collection efficiency and a low percentage of usable data.
RoboPocket is designed for large-scale distributed real-world scenario collection. Solve this impossible triangle by enhancing real-time interaction and quality control in an integrated format.
We've redefined the data collection paradigm for embodied intelligence, integrating model training understanding into edge-side intelligent hubs. Robopocket is an always-online artificial intelligence tutor. It not only instantly diagnoses the quality of each frame of data and intelligently guides collectors to adjust their actions, but also dynamically assesses the value of data through real-time interactions, ensuring that every collection directly contributes to the evolution of the models that matter most..
• Task guidance: You can send task collection tutorials to data collectors in real time to guide their operations.
Caption: Guidance for teaching assignments
• Real-time interaction reminders: Alerts you if the collector moves too fast or exceeds the robot's working space, preventing invalid data from entering the post-processing stage.
Caption: Speed abnormality detection
Caption: Abnormal operation monitoring
• Multidimensional quality scoring: Data is scored at the collection stage to allow collectors to make timely corrections and provide a basis for post-processing screening.
Caption: Closed-loop monitoring of data quality
By implementing quality control mechanisms at the collection stage, RoboPocket is able to resolve data quality issues as early as possible and move subsequent data processing from “catastrophic cleaning” to “supervised screening.”
Flexibly add first-person perspective: quickly coordinate multiple devices
Perspective limitations are a natural problem with wrist-mounted first-person views. RoboPocket supports the flexible addition of a first-person perspective and achieves the same spatial coordinates through rapid multi-device alignment.. This design maintains the “embodied coherence” of the UMI and supplements the scene context, allowing RoboPocket to not only collect “highly operational” desktop tasks, but also provide additional perspective and scene information for more complex and diverse scenarios.
Caption: first person perspective
In multi-arm acquisition or collaboration scenarios, how to quickly coordinate timestamps and unify coordinate systems for multiple devices is a key and challenging aspect. RoboPocket significantly lowers the threshold for this step. RoboPocket's fast synchronization mechanism allows multiple phones to share timestamps and SLAM coordinate systems, making pairing the two arms very easy.
Caption: High-speed synchronization of two arms
Great scalability: Rich user interactions
In addition to enhancements at the hardware level, the computing power of the iOS system and its rich UI interface bring increasingly sophisticated functional interactions, allowing data collectors to receive guidance, control quality, and even complete the entire collection process in a more intuitive manner.
In addition to the above quality control functions, Robopocket also supports real-time playback and wireless automatic uploading, and you can start collecting with voice or a button.
Data awareness, model know-how, data infrastructure
Hardware is just the starting point. The real barriers are: What is useful data? Recognizing “model know-how” and building data pipelines.
From professional data collection sites to ordinary people's daily lives, Qiongche Intelligence continues to advance the industry's understanding of data and upgrade methodologies.
• In 2023, we will release the RH20T large-scale embodied intelligence dataset in collaboration with Lu Cewu's team at Shanghai Jiao Tong University.: We were able to collect robot movement data systematically and on a large scale under preset conditions.
• In 2025, CoMiner companion data collection system released: The robot left the collection site and entered the real world to obtain richer and more complex operational data in an open environment.
• In 2026, we took a more important step: Robotic data collection from specific locations and reliance on specialized systems further released to society as a whole all normal people and all mobile phones They become nodes in the robot learning network, allowing data to continuously create value in real life.
Behind our large-scale data collection is our powerful data infrastructure and scientific data pipeline. It covers efficient task design, large-scale data collection, uploading, cleaning, quality monitoring, and finally provides model training and evaluation, providing feedback for the next round of data collection.
Caption: Toolchain built into Qiongche
The core of supporting this data and continuously creating value is the model functionality that Qiongche Intelligence has been deeply involved with for many years.
At the end of our latest public video, we demonstrated that by using only data collected by the RoboPocket solution, we can train a robot strategy that supports long-range tasks, two-arm collaboration, no teleoperation, no replay, and can execute autonomously on industrial cameras and robotic systems.
This result shows that Robopocket's data collection method, which can be put in anyone's pocket, can be stably transferred to an industrial-level recognition and execution system across the differences between collection terminals and deployment platforms, validating Qiongche Intelligence's systematic ability in data collection, data quality control, model training, and model deployment.
- Video notes: The robot autonomously performs tasks such as setting tables, clinking glasses, folding towels, and putting away snacks.
We turn “data” into real productivity. Reliably capture usable data assets from the real world, drive rapid model iterations and cost savings, deliver functionality through standardized processes to diverse scenarios like pharmacies and hotels, and enable large-scale implementation.
In the future, RoboPocket will continue to improve and enhance Qiongche's data pyramid, along with teleoperation with precision force control, CoMiner's companion field collection, and human interaction data to support the continued evolution of the embodiment model in the real world.
