Pocket Machine Learning: This time, we've packed the ability to "train robots" into everyone's pocket.

Embodied data collection is moving from collection plants to daily life

Currently, the data required for embodied intelligence is still limited to laboratory and centralized data collection scenarios. Diverse operations and physical interactions in the real world are difficult to be systematically covered. Looking back at the development of autonomous driving, we can see that when the system moved from the laboratory to real roads, data collection shifted from a small - scale, controlled approach to large - scale accumulation in real scenarios, which in turn promoted the rapid evolution of model capabilities.

Based on this experience, Qiongche Intelligence officially launched RoboPocket - through smartphones and apps, every user can become a participant in data collection, complete tasks and upload data, achieving lightweight, controllable and high - quality data collection. This extends data collection to a wider range of real environments, allowing more users to participate in task collection. On the premise of being lightweight and controllable, it continuously produces high - quality and usable data, building a more real and reliable data foundation for embodied intelligence models.

RoboPocket - enabling every ordinary person to collect high - quality data

RoboPocket is a high - quality data collection solution that every ordinary person can operate, which can be started as soon as it is taken out of the pocket.

Portable and easy to use: a highly integrated high - quality data collection device

Data collection no longer depends on complex professional equipment. A mobile phone can unlock unlimited high - quality collection. RoboPocket is lightweight and ready to use out of the box. Users only need to gently touch the phone to connect with the device and start collecting.

Caption: Unboxing and quick start

This lightweight data collection solution does not come at the cost of sacrificing accuracy.

Phones, such as the iPhone, are equipped with both RGB cameras and depth cameras (LiDAR). Compared with pure visual SLAM, it is a multi - sensor fusion (vision, depth, IMU) solution. Under this approach, the collection accuracy is higher than that of pure VSLAM. Compared with infrared positioning, there is no need to drag a base station, while maintaining a high degree of integration. The detachable and replaceable fisheye lens provides an ultra - wide field of view, and the native lens has excellent image quality, capable of withstanding challenges in real scenarios such as strong and weak light. It can maintain good performance whether in a dim room or in an outdoor environment with direct sunlight.

Caption: Fisheye lens, ultra - wide field of view

RoboPocket uses the mature capabilities of mobile phones to complete SLAM positioning and mapping. Compared with the original UMI, which requires manual calibration and repeated operations, one of the core values of RoboPocket is to make "mapping a seamless experience for users".

Caption: Rapid mapping

Ensuring collection quality in real - time, RoboPocket is also an intelligent hub

In the era of large - scale embodied models, data quality is as important as data volume.

Traditional UMI - style data collection solutions have always faced the impossible triangle of data quality, portability and ease of use, and post - processing pressure. If you want to ensure data quality, you usually need to sacrifice convenience. The collection device is connected to a computer to improve the quality feedback efficiency at the data collection site, but this method is doomed not to enter thousands of households. Or you need a small device that integrates storage and data collection, leaving the data processing pressure after collection, resulting in low collection efficiency and a low proportion of usable data.

RoboPocket is designed for large - scale, distributed real - world scenario collection. In an integrated form, it solves this impossible triangle by strengthening real - time interaction and quality control.

We have redefined the data collection paradigm for embodied intelligence, integrating our understanding of model training into the intelligent hub on the edge side. Robopocket is an always - online artificial intelligence tutor. It can not only instantly diagnose the quality of each frame of data and intelligently guide collectors to adjust their actions, but also dynamically evaluate the value of data through real - time interaction, making every collection directly contribute to the most critical model evolution.

• Task guidance: Task collection tutorials can be sent to data collectors in real - time to guide their operations.

Caption: Task teaching guidance

• Real - time interaction reminder: It reminds collectors if they are moving too fast or going beyond the robot's working space, preventing invalid data from entering the post - processing stage.

Caption: Abnormal speed detection

Caption: Abnormal movement monitoring

• Multi - dimensional quality scoring: Data is scored during the collection stage to help collectors make timely corrections and provide a basis for post - processing screening.

Caption: Closed - loop monitoring of data quality

By introducing a quality control mechanism during the collection stage, RoboPocket can solve data quality problems as early as possible, shifting subsequent data processing from "catastrophic cleaning" to "supervised screening".

Flexibly add a first - person perspective: rapid alignment of multiple devices

Limited perspective is a natural problem with the first - person wrist - mounted perspective. RoboPocket supports the flexible addition of a first - person perspective and achieves the same spatial coordinates through rapid multi - device alignment. This design retains the "embodied consistency" of UMI and supplements the scene context to ensure that RoboPocket can not only collect "highly operable" desktop tasks but also provide additional perspectives and scene information for more complex and diverse scenarios.

Caption: First - person perspective

In multi - arm collection or collaborative scenarios, the key and difficult point is how to quickly align the timestamps of multiple devices and unify the coordinate system. RoboPocket significantly lowers the threshold for this step. Through RoboPocket's rapid synchronization mechanism, multiple mobile phones can share timestamps and SLAM coordinate systems, making it extremely easy to pair two arms.

Caption: Rapid synchronization of two arms

Excellent scalability: rich user interaction

In addition to hardware - level expansion, the computing power of the iOS system and its rich UI interfaces bring more and more advanced functional interactions, enabling data collectors to receive guidance, control quality, and even complete the entire collection process in a more intuitive way.

In addition to the above quality - control functions, RoboPocket also supports real - time playback and automatic wireless upload, and collection can be started via voice or a button.

Data cognition, model know - how, and data infrastructure

The hardware is just the starting point. The real barrier lies in the cognition of what useful data is, "model know - how", and the construction of the data pipeline.

From professional data collection sites to the daily lives of ordinary people, Qiongche Intelligence continues to promote the industry's understanding of data and the upgrading of methodologies:

• In 2023, jointly released the RH20T large - scale embodied intelligence dataset with the team of Lu Cewu from Shanghai Jiao Tong University: Under preset conditions, systematic and large - scale collection of robot operation data was achieved.

• In 2025, released the CoMiner companion data collection system: Robots left the collection sites and entered the real world to obtain more abundant and complex operation data in an open environment.

• In 2026, took a more critical step: Further released robot data collection from specific places and professional - system dependence to the whole society - enabling every ordinary person and every mobile phone to become a node in the robot learning network, allowing data to continuously generate value in real life.

Behind large - scale data collection is our powerful data infrastructure and scientific data pipeline, which covers efficient task design, large - scale data collection, upload, cleaning, and quality monitoring, and finally serves model training and evaluation, and provides feedback for the next round of data collection.

Caption: Qiongche embodied toolchain

The core that supports these data to continuously create value is the model capabilities that Qiongche Intelligence has been deeply involved in for a long time.

At the end of the latest public video, we demonstrated that by using only the data collected by the RoboPocket solution, a robot strategy that supports long - range tasks, two - arm collaboration, no tele - operation, and non - replay can be trained and autonomously executed on industrial cameras and robot systems.

This result shows that the data collection method of RoboPocket, which can be "put into everyone's pocket", can cross the differences between collection terminals and deployment platforms and be stably migrated to industrial - level perception and execution systems, verifying Qiongche Intelligence's systematic capabilities in data collection, data quality management, model training, and model deployment.

• Video note: The robot autonomously performs tasks such as setting the table, clinking glasses, folding towels, and tidying up snacks.

We turn "data" into real productivity: stably obtain usable data assets from the real world, drive rapid model iteration and cost reduction, and deliver capabilities to diversified scenarios such as pharmacies and hotels through standardized processes, achieving large - scale implementation.

In the future, RoboPocket, along with high - precision force - controlled tele - operation, CoMiner companion field collection, and human operation data, will continue to improve and strengthen Qiongche's data pyramid, supporting the continuous evolution of embodied models in the real world.