Embodied AI trainingBeijing

Inside Beijing's robot school: what Shijingshan teaches and how to measure the results

Beijing's Shijingshan center runs factory and home training scenarios. Explore reported robots, data output, 4D capture and a repeatable way to test learned skills.

By TechniaHQRobot9 min read

Introduction

Beijing's Shijingshan humanoid training program has been described as a robot school because people teleoperate machines through repeated factory and household tasks. Official reporting distinguishes multiple phases of the program. A 2025 account describes a 10,000-square-meter center with 16 full-scale scenarios, while September 2026 reports describe an upgraded first-phase site of more than 1,500 square meters and separate second- and third-phase training facilities.

The reader's practical question is whether those sessions produce reliable behavior in an actual workshop. The facility's announced process connects motion capture, scene reconstruction, model training, physical validation and feedback. That process is useful only if repeated tests expose whether a trained robot can work with new objects, layouts and human interruptions.

Source review date recorded for this article November 11, 2026.

Key findings

  • The 2025 Shijingshan center announcement describes 16 life-size scenarios across manufacturing, smart homes, elder care and 5G integration; it reports a greater-than-95% success rate on practiced Kuavo tasks, without a public trial protocol that would establish field performance.
  • A September 2026 district report describes approximately 270 robot devices across the program and an annual capacity close to 20 million data items, with phase two contributing more than six million reported real-application data items per year.
  • The upgraded first-phase center adds multi-view capture and 4D scene reconstruction. Contact-force realism and autonomous operation still need direct hardware tests.

A practical scorecard for a robot trained at Shijingshan

QuestionRecordDecision supported
Does it finish the work?Complete tasks / attempted tasks, defined start and goalTask reliability
Does the skill generalize?Same tests with unseen objects and layoutsResistance to memorization
Does the robot need supervision?Interventions and recovery time per robot-hourStaffing requirement
Does the motion damage goods?Drops, collisions, torn bags, force threshold excursionsHandling quality
Can it keep pace?Cycle time, stops and charging across a full shiftProduction suitability

What is actually inside the training program?

The municipal account published in October 2025 described 10,000 square meters and 16 replicated scenarios under four categories: industrial manufacturing, smart homes, elder-care services and 5G integration. In these setups Kuavo humanoids practiced returning empty boxes, sorting materials, weighing packages and packing products with VR and motion-capture support.

On September 1, 2026, Shijingshan authorities described roughly 270 robot devices in training across a multi-phase project and annual production of close to 20 million data items. The district said phase two deployed more than 100 full-size bipedal humanoids and could produce over six million real-application data items a year. Phase three focuses on touch and multimodal sensing. These are reported program capacities, not independent counts of successful factory operations.

The September 11 account places the upgraded first-phase embodied-AI training center in the Shougang Park science-fiction innovation center, covering over 1,500 square meters. Those figures refer to different phases and reporting dates; they should not be added as though they described one room.

Problem 1: a robot repeats a package task but fails when the line changes

Repeated teleoperation can teach a robot the motions for weighing a box or placing an item on a conveyor. The main risk is overfitting to one conveyor height, box size, lighting condition or starting pose. The Beijing government itself describes loss of adaptability when an environment, device or production line changes.

A practical fix is to capture variations intentionally: change object location and orientation, introduce small geometry differences, change the route around obstacles and include both successful corrections and safe stops. Divide observations into training and unseen evaluation conditions before fitting the model.

Run a defined acceptance test on the real robot. For example, specify 30 trials each in a familiar layout and a changed layout; report full-cycle success, staff interventions, dropped packages and elapsed seconds per successful item. The number of trials here is a proposed test protocol, not a published Shijingshan result.

Problem 2: camera-realistic simulation can still mishandle contact

Shijingshan's September upgrade introduced multi-view light-field capture and a 4D world-model approach to record motion, deformation and changing scene geometry. This can give training software richer observations than a static image of a table or package.

But closing a gripper around a soft pouch depends on friction, contact force, grip compliance and the material's deformation. A scene can look correct while the simulator still predicts the wrong slip. The useful response is to measure forces and slips on hardware and use that evidence to tune the contact model.

For fabric handling or package placement, record where fingers contact the item, whether the item moves after release and whether the robot damages it. Compare the simulator's predicted outcome with the physical result under the same initial conditions.

Problem 3: more recorded data does not equal more autonomous hours

Data capacity can be quoted as frames, samples, episodes, task attempts or hours. The Shijingshan program's six-million and near-20-million figures are described as data items; without a common definition they cannot be converted into completed jobs or productive robot-hours.

Beijing's earlier municipal report said trained Kuavo robots had mastered more than 20 atomic skills and achieved over 95% success on practiced activities. The cited report does not define the full trial count, object variation, human assistance or independence of that measurement. A production site would need success rates measured over repeated shifts under normal work conditions.

To verify progress, publish a short task sheet for each skill: starting arrangement, completion rule, allowed human help, object variations, time limit, number of attempts and exact failure categories. Repeat after software changes and after a shift of continuous operation.

How the advertised training loop should close

Shijingshan describes a chain from capture to modeling to training to hardware verification and data feedback. A concrete application is a parcel handling task: record a human-guided pickup, train a policy, run the robot on new package positions, save every failed attempt with sensor timestamps and return those failures to the next data batch.

Before declaring a skill ready for a factory, measure the intervention rate per operating hour and the share of orders completed at the required speed. Factory managers can then compare robot output with a conventional fixed automation cell under the same workload and maintenance schedule.

Sources and methodology

Shijingshan humanoid training center upgradeBeijing Municipal Government / Shijingshan District · September 1, 2026The humanoid training center reopens after upgradeBeijing Municipal Government · September 11, 2026Over 10,000 square meters: humanoid robot training centerBeijing Municipal Government · October 9, 2025Shijingshan embodied-AI training upgradeBeijing Municipal Government · September 15, 2026
Article by TechniaHQRobot