Physical AI dataCross Embodiment

What Beijing's cross-embodiment data center actually records and how robots can reuse it

Inside the Haidian training center: optical tracking, glove pressure data, time alignment and the tests needed before human demonstrations work on another robot.

By TechniaHQRobot9 min read

Introduction

At the Noitom cross-embodiment data collection and training center in Beijing's Zhongguancun (Haidian) Embodied Intelligence Innovation Park, workers record motion in a near-100-square-meter optical tracking area and practice tasks in simulated homes, shops, offices and production lines. Beijing Daily's September 23 report describes the facility, and an August 24 science-commission report details the capture equipment.

The practical problem starts after the recording. A human may grasp a bottle once, yet a humanoid with shorter fingers, different motors and a different wrist cannot replay those joint angles unchanged. Teams have to clean the sensor signals, preserve the task goal, map the motion onto a new body and check the result on hardware.

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

Key findings

  • The center's fine-manipulation setup was reported to use more than 20 infrared cameras with 4-megapixel sensors at 850 nm, gloves with 3 mm tracking markers and pressure sensors. The reported sub-millimeter tracking performance describes the capture system, not robot task accuracy.
  • One Beijing Daily example records a roughly 30-second bottle handover, repeated with changes of position, direction and grip.
  • Cross-robot reuse requires data synchronization, motion retargeting and new hardware tests; there is no published success rate for universal transfer at this center.

From human motion recording to a usable robot policy

StageWhat to retainFailure to test
CaptureVideo, hand pose, pressure, joint signals, sensor clocksMissing frames or mismatched timestamps
GeometryCamera calibration, hand/object frames, robot link geometryGrip misses target object
RetargetingTask-space goal and robot joint/collision constraintsJoint saturation or unreachable pose
TrainingTrain/test split by object, layout and embodimentMemorized lab motions
Hardware testTask completion, drops, intervention, contact forceWorks in replay but fails on the robot

Problem 1: 30 seconds of video can still be unusable training data

In the September 23 facility report, a worker wearing a motion-capture suit, head tracking and sensing gloves reaches for a bottle, lifts it, turns and hands it over before releasing. The sequence takes about 30 seconds. The operator then repeats it at changed positions and angles. Recording duration alone says little about how many usable examples survive processing.

Camera frames, glove pressure, fingertip positions and the robot's own joints must describe the same instant. If clocks drift, a model may see 'contact' after the fingers have already closed; if camera-to-hand calibration is wrong, a successful reach can become a missed grasp after conversion.

An implementable pipeline is: record source clocks and calibration files; estimate the hand/object coordinate frames; align sensor timestamps; label contact onset, release and failures; reject dropped or unsynchronized frames; then split training and evaluation by object position and environment. Keep the raw take and a reason for every rejection. The center's staff describe time alignment, cleaning, labeling and quality control as required steps.

Problem 2: a human fingertip path is not a robot joint command

Beijing's science-commission account reports an optical system with over 20 four-megapixel infrared cameras, 850 nm illumination and markers on gloves. It captures finger flexion, contact position and grip force alongside first- and third-person video. Those signals are useful because the task can be described in relation to the bottle and the table instead of only as a list of human joint angles.

To retarget a demonstration, keep the goal in object coordinates, solve inverse kinematics for the target robot, constrain joint ranges and fingertip collisions, and check balance or arm reach. A three-finger gripper cannot reproduce every five-finger contact pattern. In that case the desired object motion can be preserved while the exact hand shape must change.

The remaining contact forces must be tested using the target robot's sensors. Camera capture cannot tell an operator that a different gripper is using safe forces on a deformable package. Robot-specific fine-tuning or teleoperated corrections may still be required.

Problem 3: a shared dataset can fail when the shelf, object or robot changes

In the facility's simulated supermarket, trainers use VR gear and controllers to pick goods from shelves. The August 24 report says successful grasps yield grip-force and joint-angle records. If all training examples use the same shelf height, same product and same lighting, the model may memorize that setup.

A useful transfer evaluation would hold out a robot body, some objects and a shelf layout. Train the same policy with and without human-motion data, then compare complete task success, object drops, human interventions, contact-force excursions and additional robot demonstrations required. Report trial counts and test conditions. This is a proposed evaluation design, not a result released by the Haidian center.

A published baseline shows what cross-robot pooling can and cannot establish

Google DeepMind and collaborators assembled Open X-Embodiment from more than one million real-robot trajectories covering 22 robot embodiments. Its RT-X experiments provide evidence that data collected on one platform can help training on others under defined experimental conditions. That is a useful research reference for the Beijing center.

Open X-Embodiment represents robot actions in an end-effector frame and handles missing action dimensions during training. It does not make human kinematics, robot motor torque or tactile contact identical. The Beijing facility has described its capture services; a matched cross-body benchmark from that specific facility has not been published in the sources reviewed.

What a robotics team should request before buying these data

Ask for a sample sequence that includes raw and processed signals, sensor synchronization information, coordinate transforms, robot embodiment identifiers, consent and reuse terms, and a documented train/test split. Ask how much of the captured footage is discarded during quality control.

For a chosen target hand and arm, request the number of transferred tasks completed without editing, the number of robot-specific demonstrations needed afterwards and examples of failed transfers. This makes the cost of reuse measurable against collecting fresh robot-specific demonstrations.

Sources and methodology

Inside the robot training roomBeijing Daily / Beijing Municipal Government · September 23, 2026Zhongguancun robot data training baseBeijing Municipal Science and Technology Commission · August 24, 2026Scaling up learning across many different robot typesGoogle DeepMind · October 3, 2023Robotic Learning Datasets and RT-X ModelsOpen X-Embodiment collaboration
Article by TechniaHQRobot