Introduction
A robot can record thousands of pick-and-place attempts yet produce a dataset that another team cannot train on. Missing camera calibration, unsynchronized force readings and unclear success labels can turn hours of expensive demonstrations into unreliable action targets. China's two published guidance documents address different parts of this work.
This article identifies what the official register confirms, shows a worked pick-and-place example and proposes checks an engineering team can run on its own data. The suggested tests below are an editorial engineering workflow. They are not presented as clauses, mandatory thresholds or certification requirements from the Chinese documents.
Source review date recorded for this article November 11, 2026.
Direct answer
China published GB/Z 218.1-2026, covering the quality assessment of real-world embodied AI data, and GB/Z 220-2026, covering the technical framework of data-generation platforms, on August 27, 2026. Both are national standardization guidance documents. A robotics team should examine the official texts and test timestamp alignment, coordinate frames, data provenance and task outcomes before using a dataset in training.
Published reference documents
Official standard recordsReal-world data quality
Quality assessment for real embodied AI data, including robot-collected and human demonstration data.
Data-generation platforms
A reference framework for designing and testing platforms that produce embodied AI data.
Key findings
- Both documents were issued on August 27, 2026. Their official entries identify them as current GB/Z national standardization guidance documents.
- GB/Z 218.1-2026 concerns the quality assessment of real-world data, including real-robot collection and human demonstrations; the scope does not establish a robot task success rate.
- GB/Z 220-2026 describes a platform framework with engine, generation and application layers. Its focus is platform design, development and testing.
- A dataset audit still needs task-specific checks: clock alignment, coordinate frames, calibrated signals, failure labels and held-out robot trials.
Fields that make robot datasets easier to combine
| Area | Example field | Reason |
|---|---|---|
| Time | Sensor timestamp and clock source | Aligns images, joints and force |
| Geometry | Coordinate frame and calibration | Makes trajectories interpretable |
| Task | Start state, goal and success rule | Prevents ambiguous labels |
| Assistance | Human intervention flag | Separates autonomous from assisted data |
| Provenance | Device, site and collection method | Supports auditing and reuse |
| Safety | Contact or stop event | Preserves failure evidence |
The two documents published in August 2026
The State Administration for Market Regulation and the Standardization Administration of China published both entries on August 27, 2026. The official standards register lists GB/Z 218.1-2026 as 'Artificial intelligence—Specifications for embodied intelligence data quality—Part 1: Real data' and GB/Z 220-2026 as 'Artificial intelligence—Technical requirements for embodied intelligence data generation platform'.
The GB/Z designation identifies national standardization guidance documents. A published guidance document is a useful technical reference; whether a specific contract, procurement tender or regulation requires compliance must be checked separately. The database entry alone cannot establish a legal obligation for every robot manufacturer.
The first document concerns the quality of real data. The second describes a reference framework for generating data. They address different review questions for a team building a training pipeline.
Published documents and their stated scope
| Document | Official subject | Practical question | Status |
|---|---|---|---|
| GB/Z 218.1-2026 | Quality of real embodied AI data | Can recorded trajectories be assessed for training use? | Current guidance · 27 Aug 2026 |
| GB/Z 220-2026 | Embodied AI data-generation platforms | How should a platform be organized and tested? | Current guidance · 27 Aug 2026 |
Titles, identifiers, issue date and status are verified in the Chinese national standards register. The practical questions are TechniaHQRobot's interpretation, not verbatim requirements.
The first failure: camera and force sensors record different moments
Imagine a gripper reaching toward a bottle. The RGB camera records the fingers approaching, the force sensor reports contact and joint encoders report the closure. The dataset may store each stream at a different sampling frequency. If one clock runs late or a frame drops, the data can pair a closed gripper with an image from before contact.
Before training, record the clock source for every sensor, the acquisition timestamp, the time offset estimated during synchronization and any missing samples. Replay the streams together around contact events. Mark gaps instead of filling them silently when the missing interval includes a grasp or slip.
An engineer can create a timing test with a visible contact event and a force spike, then compare their recorded timestamps. Record the observed offset and repeat after long sessions. The offset tolerance depends on the controller rate and the task; this article does not assign an arbitrary threshold to the Chinese standard.
Audit question: Can a reviewer reconstruct the instant of first contact from the camera, joint state and force stream without guessing which sample came first?
The second failure: a recorded action has no reliable coordinate frame
A camera can locate a bottle relative to its optical frame, while a robot arm executes moves relative to its base. A tool pose stored without the camera-to-base transform has no unambiguous physical meaning. A recorded hand path may look sensible in a visualization and still miss the bottle when replayed on hardware.
Keep the camera intrinsics, sensor extrinsics, robot URDF or geometry version, joint units and a description of the coordinate convention. During collection, run a calibration check on known points in the work area and log the measured position error. Store which calibration file was active for each demonstration.
When data moves between robots, convert the object-relative task goal and solve for the target robot's joint limits and reach. A recorded human wrist angle cannot be treated as an executable robot command without retargeting and a physical validation test.
Worked example: what one pick-and-place episode should contain
Consider a tabletop task: the robot picks up a 250 g bottle and places it in a marked tray. The mass here is an illustrative test condition, not a specification of either Chinese document. A useful episode records the initial bottle pose, the intended tray position, all sensor timestamps, the robot action sequence, the first contact, the release and the final outcome.
The final label should state whether the bottle reached the tray without falling, whether an operator intervened and whether the object was damaged. Include an unsuccessful reach and a recovery attempt in the dataset instead of discarding them automatically. A failure example can help a team study what the controller should avoid and how it should recover.
A training file alone cannot prove a policy works. After training, run the robot with new bottle positions and tray locations that were excluded from training, then count complete placements. Keep the same evaluation protocol when comparing versions.
An example record for one bottle-moving task
| Field | Example value or format | Why it is needed |
|---|---|---|
| episode_id | Unique identifier | Links images, sensors, actions and the result |
| clock and timestamp | Clock ID plus per-sample time | Reconstructs sensor order around contact |
| calibration_id | Camera and robot frame calibration version | Converts observations into the robot frame |
| robot and gripper | Model, firmware and tool configuration | Prevents mixing incompatible actuators |
| task start and goal | Object pose, tray pose and allowed area | Defines the intended physical outcome |
| action and contact | Joint state, commanded action, force samples | Explains grasp and release behavior |
| result and intervention | Completed / dropped / assisted / damaged | Preserves trial outcome and failure details |
This is an illustrative TechniaHQRobot dataset schema for readers. The listed fields are not presented as quoted mandatory fields from GB/Z 218.1-2026.
Which checks catch unusable training trajectories?
A practical audit should examine a sample of episodes from the start, middle and end of every collection session. Reject samples that refer to a missing robot configuration, have invalid timestamps or cannot be aligned with the relevant action. Log whether an episode was excluded and why, so a second team can reproduce the selection.
Start with three simple measurements: the share of samples with complete required fields, the share with usable calibration and the share whose result labels pass manual review. Break these rates down by robot, task, recording site and collection date. A single overall quality score can hide one malfunctioning camera or a misconfigured glove.
Keep contact failures, interrupted tasks and other rare events when they have valid sensor records. They are useful for evaluation and recovery learning, although they need explicit labels so they are not mistaken for successful demonstrations.
Problem, engineering check and what should be recorded
| Observed problem | Check | Evidence to retain |
|---|---|---|
| Force reading arrives after the grasp | Compare synchronized contact event timestamps | Offset distribution and clock source |
| Robot misses a previously recorded target | Replay calibration targets and check transforms | Calibration version and position error |
| Data includes unreported operator help | Review intervention flags against session logs | Assistance marker and episode ID |
| Only repeated easy examples remain | Inspect distribution by object, task and failure | Split counts and sampling criteria |
| A new robot performs worse | Evaluate on withheld hardware and placements | Full task outcomes and trial counts |
How the data-generation platform fits into the workflow
GB/Z 220-2026 describes a reference framework for embodied AI data-generation platforms and names three layers: engine, generation and application. Its stated use covers the design, development and testing of these platforms. The public scope statement is sufficient to identify the layers, although this article does not reproduce the standard's full clauses.
For a sample engineering architecture, the engine layer can provide scene dynamics, robot and sensor interfaces, while the generation layer produces and varies task episodes. The application layer can export datasets and report coverage to training teams. This mapping is an explanatory design example; it must be checked against the licensed standard before it is treated as a conformity claim.
If a team augments recorded trajectories with simulated ones, it should keep a provenance flag for real, human-demonstrated and generated samples. Evaluate the resulting policy on physical hardware, because convincing rendered scenes do not establish accurate friction, flexible-material behavior or safe contact forces.
How to compare real recordings, human demonstrations and synthetic data
Real-robot data includes the control commands, delays and contact behavior of the actual machine. Human demonstrations can cover a range of practical tasks but often require a mapping from human motion to the target robot. Simulation can vary scenes and objects at scale, provided the generated labels and physics are checked.
Use a training comparison that holds the robot, compute budget and evaluation tasks fixed. Report the number of real episodes, the number of synthetic episodes and the task completion rates on test objects unseen during training. Without that breakdown, a larger dataset may appear better because it contains more repetitions of an already-solved task.
Preserve a record of what was generated and what was physically observed. A synthetic contact force should not be stored as though a force sensor measured it on a real gripper.
Different data sources and their validation needs
| Source | What it captures | Check before training |
|---|---|---|
| Real robot | Actions and contact on the target hardware | Calibration, failures, intervention and clock integrity |
| Human demonstration | Human motion and task strategy | Retargeting, contact semantics and consent |
| Simulation or augmentation | Programmatically varied environments and labels | Physics mismatch and real-robot validation |
A practical checklist before using a robot dataset
The checks below are intended for teams comparing a new dataset or planning a collection contract. Use thresholds that fit the machine, task speed and safety assessment; neither document's official register entry supplies a universal success-rate threshold.
A useful acceptance exercise takes a frozen dataset, a documented policy and a held-out set of real tasks. Run repeated trials with changed object poses, count complete task success and separate sensor faults from planning, grasp and placement failures. Publish the test conditions beside the numbers.
The national standards register also lists a separate November 2026 drafting project for industrial high-quality embodied AI datasets. It is still marked as being drafted. Do not cite that project as a finished document.
- Verify the official identifiers and obtain the full texts if your team needs a formal clause-by-clause review.
- Inventory data sources, robot models, camera and joint rates, coordinate frames and calibration versions.
- Check timestamps around contact, record missing samples and validate any resampling method.
- Review task labels, failures, human assistance and the origin of every synthetic episode.
- Split the dataset by task conditions before training and evaluate on real hardware that was held out.
- Keep trial counts, contact or damage events, model versions and test dates with the results.
Sources and methodology
Document identifiers, English titles, issue dates and current status were checked against the Chinese national standards register and the 2026 publication notice on November 11, 2026. Public scope summaries were used for the distinction between real-data quality and data-generation platforms.
Worked examples, field names, suggested audits and the bottle-moving scenario are TechniaHQRobot engineering illustrations. They are not copied clauses or formal conformity checks. For contractual or regulatory interpretation, obtain and review the full official documents.