Embodied Foundation Models After Regulatory Filing: Compliance and Engineering Follow-Up
This article explains the compliance and engineering work that follows regulatory filing for embodied foundation models. It covers scoping applicable reviews, assembling filing evidence, separating content safety from functional safety, implementing a layered safety architecture with emergency stops, staged rollout and rollback, an implementation checklist, and investment priorities.
First Determine Which Reviews Apply
Embodied foundation models do not only output text; they also drive robot actions, so compliance requirements are more complex than those for pure digital services. Obtaining a filing only means one gate on the content side has been cleared. Teams need to answer the following questions first to determine the direction of material preparation.
Is the product offered to the public in China?
Example: Voice Q&A, companion dialogue, and household command understanding for home users usually count as public-facing services and require filing under the Interim Measures for the Administration of Generative Artificial Intelligence Services. For tools used only for internal production-line debugging, whether they apply still needs to be confirmed according to the competent authority’s position.Does the product have public opinion attributes or social mobilization capacity?
Example: Products with public posting, comments, community interaction, or group dissemination features need to be assessed item by item. Devices that operate purely locally usually carry lower risk, but the assessment record should still be retained.Does the product include algorithmic recommendation?
Example: Recommending service modes based on user history, pushing task reminders, or displaying content ranking all count as algorithmic recommendation. If the conditions are met, algorithm filing should be completed under the Administrative Provisions on Algorithm Recommendation for Internet Information Services.Does the data involve cross-border transfer or important data processing?
Example: Sending home-environment video back to an overseas cloud for training is a cross-border data transfer and may require a security assessment filing in accordance with the law. Processing process parameters on a production line requires confirmation of whether the data counts as important data.Does the product process sensitive personal information?
Example: Home camera images, voiceprints, faces, addresses, and child-related information are all sensitive personal information. Processing such information requires notice and consent, minimum necessity, and protection impact assessment under the Personal Information Protection Law and the Data Security Law.
The above judgments are subject to the competent authority’s actual interpretation. The same product may involve multiple reviews at the same time. It is recommended that legal and compliance personnel form a written conclusion at the project initiation stage and retain the basis for the judgment.
Filing Evidence Checklist
Filing review usually focuses on the content controllability of the model. Teams should organize the following evidence in advance, and each item needs an owner and a storage location.
| Evidence item | Example evidence | Owner | Retention requirement |
|---|---|---|---|
| Corpus source and authorization chain | Data source ledger, authorization agreements, copyright review records | Data lead, legal | Archive with each training version |
| Content safety evaluation | Test set description, violation content interception rate, test date and model version | Algorithm lead | Re-evaluate after each model update |
| Annotation quality | Annotation guidelines, spot-check records, annotator consistency metrics | Data annotation manager | Keep spot-check samples and conclusions |
| Refusal capability | Refusal test case set, refusal rate, false refusal rate | Algorithm lead, product manager | Archive together with content safety evaluation |
| Red-team test records | Test scope, issues found, remediation and retest results | Security engineer | Record remediation closure; do not record reusable attack details |
Embodied scenarios also need one more evidence category: action command out-of-range testing. The test checks whether the system intercepts action requests when model output exceeds the preset range, and records the interception count and handling results. These records support both filing review and subsequent functional safety assessment.
Example threshold: The target content interception rate is set by the team based on scenario risk. All numeric values in this article must be re-determined for the specific robot model and scenario.
Separate Content Safety from Functional Safety
Filing addresses “what the model says.” Robot action safety addresses “what the robot does.” They cannot substitute for each other.
The output mode of an embodied model determines the risk scope. End-to-end models directly output joint or end-effector commands and carry the highest risk. In layered architectures, the large model only outputs intent or skill calls, and the planning module, motion controller, and safety module check layer by layer, giving better controllability. Not all embodied models directly drive actuators. Teams should first confirm which type they are before deciding the focus of safety investment.
Layered Safety Architecture and Emergency Stop
The recommended layered structure is as follows, and each layer needs clear checks.
- Model layer: Outputs skill names and parameters, not raw motor commands. The check is whether the output format is constrained and whether invalid formats are discarded.
- Planning and validation layer: Maps commands to a predefined skill library and checks whether parameters are within the valid range. The check is whether the skill library version matches the model version.
- Motion limiting layer: Limits speed, acceleration, torque, and joint angle range.
Example value: Maximum end-effector speed of 0.5 m/s. Measurement basis: Instantaneous speed of the end-effector relative to the robot base coordinate frame; determine after actual measurement on the robot model. - Motion control and safety control layer: Runs independently of the large model and is responsible for collision detection, torque limiting, and safety stop. The code and configuration of this layer should be version-controlled separately.
- Emergency stop layer: The hardware E-stop circuit is independent of software and, when triggered, cuts power or enters a safe stop state. Software E-stop is only supplementary.
Example threshold: No more than 200 ms from E-stop trigger to the robot entering a safe state. Measurement basis: Time from the trigger signal to joint speed reaching zero; it must be measured on the target robot model and the values in this article cannot be directly reused.
Least privilege is equally important. Operations, maintenance, and development accounts should be separated, and each should access only the necessary interfaces. Manual switching and logging are required between maintenance mode and operation mode to avoid accidentally entering a debug state in user scenarios.
Risk note: Robots in home and production-line scenarios need a physical safety assessment, including hazard identification, human-robot interaction distance, and consequences of erroneous actions. Passing filing does not mean functional safety is met. Teams should conduct assessments with reference to applicable robot safety standards and retain the assessment report.
Staged Rollout and Rollback
Release should proceed in stages, and each stage needs entry conditions.
- Stage 1: Simulation environment regression testing, covering all skills and out-of-range command cases.
- Stage 2: Closed-site real-robot testing, focusing on E-stop and limiting.
- Stage 3: A small controlled user group, with a manual takeover channel.
- Stage 4: Gradually expand the scope, and review metrics before each expansion.
Example gate: No serious safety events in N consecutive tests; N is determined by risk level. Gate values should be signed off by the safety lead.
Suggested metrics include refusal rate, motion limiting trigger count, E-stop count, manual takeover rate, and user complaint categories. If metrics are abnormal, pause expansion and investigate the cause first.
The rollback mechanism needs to be prepared in advance. Retain the previous validated model version, data version, and safety parameter configuration, and ensure one-to-one correspondence among the three. Rollback operations should be written as an executable step list. Example target: Complete rollback within 30 minutes; the specific time limit is determined by the operational service-level agreement. Rollback drills should be conducted regularly, and drill records should be retained.
Implementation Checklist
Filing materials
- Applicability scope has been determined, and a written conclusion has been retained.
- The corpus ledger, authorization agreements, and copyright review records have been archived.
- The content safety evaluation report matches the current model version.
- The refusal test case set and false refusal rate statistics are complete.
Algorithm and safety assessment
- Whether algorithmic recommendation is included has been determined, and the conclusion is recorded.
- Whether cross-border data transfer or important data is involved has been determined.
- Safety assessment materials have been prepared according to competent authority requirements.
Data and personal information
- The notice and consent process for sensitive personal information has been launched.
- The collection scope meets the minimum necessity principle.
- The protection impact assessment report has been completed.
Engineering release control
- The layered safety architecture has been implemented, and large-model output does not directly drive actuators.
- Motion limiting parameters have been measured on the target robot model, with the measurement basis documented.
- The hardware E-stop has been tested, and response time is recorded.
- Entry gates for each staged rollout have been signed by responsible persons.
- Rollback steps have been drilled, and drill records have been retained.
Investment Priorities
When resources are limited, invest in the following order. First, complete the applicability judgment and evidence ledger; this is the prerequisite for filing and the foundation for all subsequent work. Second, prioritize hardware E-stop, motion limiting, and permission isolation; these directly affect personal and property safety and should be completed before launching any user scenario. Third, establish staged rollout and rollback processes so that losses can be contained quickly when issues occur. Fourth, continuously improve evaluation sets and red-team testing, and iterate with model updates.
Filing is only the starting point for compliance. After launch, the model, data, and scenarios will change, and evidence and safety parameters need to be updated in sync.
