A method and system for safety scheduling of a humanoid robot with risk time delay

Abstract

This invention discloses a risk-latency-safe scheduling method and system for embodied robots. The method acquires the robot's multimodal state, task instructions, and edge resource state to generate action segment scheduling descriptors. It calculates a risk-latency joint score by comprehensively considering task risk, model confidence, sensor freshness, latency budget, edge load, and the degree of action irreversibility. Based on this, it dynamically selects the inference position and model specifications, and determines the action length, execution frequency, and degradation strategy. Candidate actions undergo dual semantic and physical safety reviews to generate safety leases. These leases limit the effective time and execution boundaries, automatically expiring and triggering a safety stop or takeover when environmental changes, sensor expiration, inference timeout, or increased risk occur. This invention unifies edge scheduling and safety governance into a closed-loop mechanism, improving action safety, auditability, and continuous operation under network anomalies while reducing inference latency.

Description

Technical Field

[0001] This invention belongs to the field of embodied intelligent robot technology, specifically relating to a method and system for risk-delay safety scheduling of embodied robots. Background Technology

[0002] Embossed intelligent robots typically understand tasks through multimodal information such as vision, language, touch, force, odometry, and joint states, and generate executable actions through vision-language-action models or basic robot models. With the development of large-scale visual language models and robot strategy models, robots are able to perform complex tasks such as grasping, moving, opening doors, assembling, sorting, delivering, and inspecting according to natural language instructions.

[0003] In existing technologies, the deployment of embodied intelligent robot models typically employs one or a combination of ontology-based inference, edge inference, or cloud inference. Ontology-based inference offers advantages in low communication latency and privacy, but is limited by computing power, power consumption, and heat dissipation. Edge inference provides strong computing power and low local area network latency, but is affected by edge node load, network jitter, and concurrent tasks of multiple robots. Cloud inference offers stronger model capabilities, but carries higher risks in communication latency, network availability, and data security. Existing technologies primarily optimize latency through model compression, asynchronous inference, pipelined execution, or cloud-edge collaboration.

[0004] On the other hand, the safety issues of embodied intelligent robots not only include collisions, speed, torque, and workspace constraints in traditional robot control, but also behavioral risks caused by natural language instructions, scene misunderstandings, model illusions, semantic attacks, access violations, and long-term planning deviations. Existing technologies typically add safety filters before model input or before action execution, or reduce dangerous actions by training models with safety preferences.

[0005] However, existing solutions often separate real-time scheduling and security governance: the scheduling module primarily focuses on inference throughput and latency, while the security module primarily focuses on whether actions violate regulations. For embodied intelligent robots, the safety of the same action segment is closely related to its inference latency, sensor freshness, the rate of environmental change, the irreversibility of the action, available computing resources, and execution window. For example, when a person approaches, an obstacle moves rapidly, or a clamping action is about to close, even if the model output itself is semantically reasonable, risks may arise due to outdated inference results or excessively long action windows. Therefore, a closed-loop runtime mechanism that can unify edge resource scheduling, action latency, semantic safety, physical safety, and degrade takeover is needed. Summary of the Invention

[0006] To address the aforementioned problems in the existing technology, the present invention aims to provide a method and system for risk-delay-safe scheduling of embodied robots.

[0007] This invention provides the following technical solution: a risk-delay safety scheduling method for embodied robots, comprising:

[0008] S1. Obtain robot multimodal state, user task instructions, environmental semantic information, edge computing resource state, network state, and candidate action fragments generated by vision-language-action model inference;

[0009] S2. Generate scheduling descriptors based on candidate action fragments and calculate the risk-latency joint score;

[0010] S3. Based on the risk-latency joint score, dynamically determine the inference deployment location, model specifications, and action execution strategy;

[0011] S4. Conduct semantic security governance review and physical security constraint review of candidate action segments;

[0012] S5. When a candidate action fragment simultaneously satisfies the scheduling condition, semantic safety condition, and physical safety condition, a security lease is generated.

[0013] S6. Execute the candidate action fragment during the validity of the safety lease;

[0014] S7. When a security lease expires or is revoked, trigger the corresponding degradation strategy based on the reason for the expiration.

[0015] Furthermore, in step S2, the process of generating the action segment scheduling descriptor is as follows:

[0016] 1) Extract the task semantics, candidate action trajectory, action fragment length, and safety constraints of candidate action fragments;

[0017] 2) Based on the information extracted in step 1), associate the current sensor timestamp, model confidence, inference time, resource requirements, and network status;

[0018] 3) Based on the information associated in step 2), combined with the risk level and available degradation strategies, generate an action segment scheduling descriptor containing the expected execution window.

[0019] Furthermore, in S2, the calculation process for the risk-delay joint score is as follows:

[0020] 1) Quantify the task risk, model confidence, and degree of irreversibility of the action;

[0021] 2) Obtain sensor freshness, edge computing load, distance between people and obstacles, and network status;

[0022] 3) Based on the end-to-end latency budget, the risk-latency joint score is calculated according to the quantification results in step 1) and the state parameters obtained in step 2).

[0023] 4. The method according to claim 1, wherein the specific process of step S3 is as follows:

[0024] S3.1 Based on the risk-latency joint score, select the robot body, edge, or cloud as the inference deployment location;

[0025] S3.2 Based on the inference deployment location and task requirements, select a lightweight model, standard model, or high-precision model as the model specification;

[0026] S3.3 Based on the risk-delay joint score and delay budget, determine the action segment length, execution frequency, and scheduling priority.

[0027] Furthermore, in S4, the semantic security governance review process includes:

[0028] 1) Parse and validate user commands, task semantics, scenario semantics, permission rules, and behavioral norms;

[0029] 2) Based on the parsing and verification results of step 1), identify dangerous instructions, unauthorized instructions, ambiguous instructions, privacy risks, access to prohibited areas, abnormal interaction objects, and behaviors inconsistent with the task objectives; among them, abnormal interaction objects include objects that do not match the task objectives, user permissions, object categories, or security rules;

[0030] 3) Based on the recognition results of step 2), determine whether the candidate action meets the semantic safety requirements.

[0031] Furthermore, the physical security constraint review process in S4 includes:

[0032] 1) Perform collision prediction, workspace constraints, speed constraints, acceleration constraints, torque constraints, contact force constraints, and human-robot distance constraints on candidate motion segments;

[0033] 2) Based on the constraints in step 1), perform control barrier function constraints and action reversibility checks;

[0034] 3) Based on the inspection results of step 2), output the physical security determination result.

[0035] Furthermore, in S5, the process of generating a security lease includes:

[0036] 1) The judgment results of the aggregation scheduling conditions, semantic security conditions, and physical security conditions are used to obtain a comprehensive judgment result;

[0037] 2) Based on the comprehensive judgment results, generate authorization information including the effective time window, allowable state deviation, allowable risk level, and action boundaries;

[0038] 3) Based on the authorization information, configure the execution frequency, degradation strategy and revocation conditions to form an executable security lease.

[0039] Furthermore, the description also includes:

[0040] S8. Based on the hybrid critical level, the robot's operation tasks are divided into C0 level tasks, C1 level tasks, C2 level tasks and C3 level tasks;

[0041] For C0-level tasks, it runs independently of the vision-language-action model and performs emergency stops, collision constraints, speed limits, torque limits, and safety boundary monitoring in real time.

[0042] For C1 level tasks, local state estimation, trajectory tracking, and low-level control are performed;

[0043] For C2 level tasks, perform visual-language-action model inference, action fragment generation, and short-range task execution;

[0044] For C3 level tasks, perform long-term task planning, cloud-based large model inference, log analysis, and strategy updates.

[0045] Furthermore, it also includes:

[0046] S9. When the network is interrupted or the edge is unavailable, cloud inference or edge inference is prohibited, and the robot body's lightweight model, conservative skill library and C0 level safety control are switched.

[0047] S10. Record scheduling decisions, security reviews, lease issuance, lease cancellation, action interception, degraded execution, manual takeover, and near-miss events;

[0048] S11. Based on the records in S10, update the task risk profile, model delay profile, scenario safety rules, and red team test cases.

[0049] A risk-delay safety scheduling system for embodied robots includes:

[0050] The status acquisition module is used to acquire the robot's real-time perception information and task semantic information;

[0051] The VLA inference module is used to generate candidate action fragments;

[0052] The edge resource scheduling module is used to generate action fragment scheduling descriptors and calculate the risk-latency joint score;

[0053] The semantic security governance module is used to perform semantic security reviews on candidate action fragments;

[0054] The physical safety constraint module is used to perform physical safety reviews on candidate action segments;

[0055] The security lease management module is used to generate a security lease when a candidate action fragment passes the review, and to revoke the security lease when the revocation conditions are met.

[0056] The execution control module is used to execute action segments during the validity of the security lease and to trigger security degradation operations when the security lease expires or is revoked.

[0057] The audit update module is used to record runtime events and update risk profiles and delay profiles;

[0058] The hybrid critical-level safety control module operates independently of the VLA inference module and is used to perform emergency stops, collision constraints, speed limits, torque limits, and safety boundary monitoring.

[0059] By employing the above-described technology, the beneficial effects of the present invention compared to the prior art are as follows:

[0060] 1) This invention incorporates risk-latency joint scoring calculations by combining risk-related parameters such as task risk, model confidence, sensor freshness, and irreversibility of actions with resource-related parameters such as end-to-end latency budget, edge computing load, and network status. This ensures that inference location selection, model specification matching, action segment length setting, and degradation strategy triggering are simultaneously subject to resource constraints and security risk constraints, thus solving the problem that traditional scheduling only focuses on computing power and latency and is disconnected from security mechanisms.

[0061] 2) This invention uses a security lease to bind constraints such as an effective time window, allowed state deviation, action boundary, and cancellation condition to the action segments output by the VLA model. When the sensor expires, the environmental change exceeds the threshold, the inference timeout occurs, or the execution deviation exceeds the limit, the lease is automatically cancelled and the action execution is blocked, thus preventing the static inference results from being continuously executed in the dynamic environment.

[0062] 3) This invention defines emergency stop, collision constraint, speed limit, torque limit and safety boundary monitoring as C0 level tasks by using a hybrid critical level division. C0 level tasks run independently of the vision-language-action model and are not affected by the inference state of the large model or the network connection state, thus ensuring the availability of low-level safety control under extreme conditions.

[0063] 4) Under the guidance of risk-latency joint scoring, this invention supports the dynamic selection of inference location between the ontology, edge, and cloud, and matches different specifications of lightweight model, standard model, and high-precision model; when the network is interrupted or the edge is unavailable, it automatically switches to the lightweight model, conservative skill library and C0 level security control on the ontology to maintain the basic task execution capability.

[0064] 5) This invention records scheduling decisions, security review results, lease status changes, action interception events, downgrade execution strategies, manual takeover records, and security near-miss events throughout the entire process through an audit update module. Based on the records, the task risk profile, model delay profile, and scenario security rules can be iteratively updated, which facilitates subsequent policy updates, incident analysis, and compliance assessment. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of the architecture of the embodied robot risk delay safety scheduling system in an embodiment of the present invention;

[0066] Figure 2 This is a schematic diagram of the risk-delay safety scheduling method for embodied robots provided in an embodiment of the present invention;

[0067] Figure 3 A schematic diagram of the security lease lifecycle provided for embodiments of the present invention;

[0068] Figure 4 This is a schematic diagram of the risk-delay joint scheduling strategy provided in an embodiment of the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0070] Conversely, this invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined in the claims. Furthermore, to provide a better understanding of the invention, certain specific details are described in detail below. However, those skilled in the art will fully understand the invention even without these detailed descriptions.

[0071] Example:

[0072] This embodiment provides a method for a risk-delay safety scheduling system for embodied robots. The system is based on a system implementation that includes a state acquisition module, a VLA inference module, an edge resource scheduling module, a semantic safety governance module, a physical safety constraint module, a safety lease management module, an execution control module, a hybrid critical-level safety control module, and an audit update module. The specific steps of the method are as follows:

[0073] Step S1: Obtain multimodal states, task instructions, and candidate action fragments

[0074] The status acquisition module collects the robot's own status and the environment's status.

[0075] The robot's physical state includes the joint angles of the robotic arm, joint speeds, joint torques, end effector pose, gripper opening and closing state, chassis pose, battery state, controller state, and emergency stop state.

[0076] The environmental conditions include RGB images, depth images, point clouds, obstacle locations, human locations, target object locations, tactile array data, contact force data, and semantic maps.

[0077] The VLA inference module generates candidate action fragments based on user task instructions and multimodal states.

[0078] Candidate motion segments can be end-effector pose sequences, joint trajectories, velocity commands, gripper control commands, chassis motion commands, discrete skill call sequences, or combinations of the above commands.

[0079] The VLA inference module can be deployed on the robot itself, at the edge, or in the cloud, and supports lightweight models, standard models, and high-precision models.

[0080] Step S2: Generate scheduling descriptors and calculate the risk-latency joint score.

[0081] The edge resource scheduling module generates action segment scheduling descriptors for candidate action segments.

[0082] The scheduling descriptor includes at least the task semantics, candidate action trajectories, expected execution window, action fragment length, sensor timestamps, model confidence, inference time, resource requirements, network status, risk level, security constraints, and available degradation strategies.

[0083] The edge resource scheduling module maintains resource profiles for the on-device, edge, and cloud, including processor utilization, graphics processor utilization, neural network accelerator utilization, video memory usage, queue length, average inference time, inference time quantile, network round-trip latency, packet loss rate, and number of concurrent tasks.

[0084] The edge resource scheduling module also maintains a model latency profile to predict the completion time of a specified model on a specified node.

[0085] Risk-delay joint scoring can be achieved through weighted rules, decision trees, Bayesian models, reinforcement learning strategies, or rule-based state machines.

[0086] Scoring inputs include, but are not limited to, task category, action category, model confidence, sensor time difference, minimum distance of predicted trajectory, edge load, network round-trip latency, degree of action irreversibility, and frequency of historical near miss events.

[0087] The score output is used to select the inference location, model size, action segment length, and safety lease parameters.

[0088] Step S3: Determine the inference deployment location, model specifications, and action execution strategy.

[0089] Based on the risk-latency joint score, select the robot body, edge, or cloud as the inference deployment location, select a lightweight model, standard model, or high-precision model as the model specification, and determine the action segment length, execution frequency, and scheduling priority.

[0090] In one implementation, lightweight models, standard models, and high-precision models are categorized according to model specifications under the same test task set and operating environment. These specifications include inference latency per action segment, memory usage, computational load per inference segment, trajectory position error, attitude error, task success rate, and action safety compliance rate.

[0091] The lightweight model is one with a single action segment inference latency of no more than 100ms, a video memory usage of no more than 2GB, a single inference computation of no more than 20GFLOPs, a trajectory position error of no more than 50mm, an attitude error of no more than 10 degrees, a task success rate of no less than 80%, and an action safety compliance rate of no less than 90%.

[0092] The standard model is one with a single action segment inference latency of no more than 300ms, a video memory usage of no more than 8GB, a single inference computation of no more than 100GFLOPs, a trajectory position error of no more than 25mm, an attitude error of no more than 5 degrees, a task success rate of no less than 90%, and an action safety compliance rate of no less than 95%.

[0093] The high-precision model is one with a single action segment inference latency of no more than 800ms, a video memory usage of no more than 24GB, a single inference computation of no more than 500GFLOPs, a trajectory position error of no more than 10mm, an attitude error of no more than 2 degrees, a task success rate of no less than 95%, and an action safety compliance rate of no less than 98%.

[0094] The above indicators can be preset or adjusted according to robot type, task category, execution scenario and deployment hardware.

[0095] The risk-delay joint scheduling strategy includes at least one of the following operating modes:

[0096] 1) Low-risk, low-latency mode: Prioritize lightweight models on the main body or edge, and use longer action segments to improve execution continuity;

[0097] 2) High-risk, low-latency mode: Employs short action segments, high execution frequency, and strict security leases, and prioritizes inference on the main body or low-load edge.

[0098] 3) Low-risk, high-latency mode: Long-term planning or non-real-time semantic analysis is handled by the cloud, while the ontology maintains local security control.

[0099] 4) High-risk, high-latency mode: Prohibit the execution of highly irreversible actions, and instead execute safe stop, conservative skill library actions, or request manual takeover.

[0100] When the risk-delay joint score exceeds the first threshold, shorten the action segment length and increase the lease renewal frequency;

[0101] When the risk-delay joint score exceeds the second threshold, switch to the conservative skills pool or request manual confirmation.

[0102] When the risk-delay joint score exceeds the third threshold, the execution of candidate action segments is prohibited.

[0103] A safety lease can include the maximum number of execution steps for an action segment. If the VLA model outputs an action sequence of twenty control steps, and the current risk level is high, the action segment will be truncated into five control steps, and a new lease will be required every five control steps.

[0104] Step S4: Semantic security governance review and physical security constraint review

[0105] The semantic security governance module reviews natural language instructions, scene semantics, and candidate action semantics.

[0106] The review rules include prohibitions against harming persons, damaging property, entering restricted areas, operating unauthorized equipment, handling dangerous items, disclosing privacy information, performing operations that are incompatible with the user's role and permissions, and performing actions that conflict with the mission objectives.

[0107] The semantic security governance module converts rules into action preconditions, temporal logic constraints, object permission constraints, or task state machine constraints.

[0108] The physical safety constraint module performs a physical safety review on candidate actions.

[0109] Physical safety review includes collision detection, minimum distance prediction, velocity constraints, acceleration constraints, torque constraints, contact force constraints, workspace constraints, stability constraints, control barrier function constraints, and action reversibility constraints.

[0110] When the predicted trajectory of a candidate action conflicts with a human body, obstacle, restricted area, or robot self-collision area, the physical safety constraint module outputs a rejection result or a corrected action boundary.

[0111] Step S5: Generate a secure lease

[0112] When a candidate action fragment simultaneously meets the scheduling conditions, semantic safety conditions, and physical safety conditions, the security lease management module generates a security lease.

[0113] The safety lease includes lease identifier, robot identifier, task identifier, motion segment identifier, effective start time, effective end time, permissible pose deviation, permissible speed deviation, permissible risk level, permissible object to be operated on, permissible workspace, permissible contact force range, execution frequency, cancellation conditions, and degradation strategy.

[0114] Security leases can be implemented as data structures in edge runtime or encapsulated as execution authorization tokens sent to the security controller.

[0115] The conditions for revocation include sensor timestamps exceeding the freshness threshold, environmental changes exceeding the change threshold, model inference exceeding the latency budget, action execution deviation exceeding the allowable deviation, personnel or obstacles entering the safety boundary, changes in task semantics, and the safety controller triggering an alarm.

[0116] Security leases go through the following stages in sequence: application, review, issuance, execution, renewal, cancellation, and filing.

[0117] Once the candidate action fragment is generated, the execution control module submits a lease application to the security lease management module.

[0118] The security lease management module calls upon the semantic security governance module and the physical security constraint module for review. Once the review is passed, the lease enters the issuance state.

[0119] During execution, if the action segment is about to be completed and the environment changes little, the lease can be renewed based on the new sensor status.

[0120] If the sensor timestamp exceeds the freshness threshold, the action execution deviation exceeds the allowable deviation, the model inference exceeds the latency budget, personnel or obstacles enter the safety boundary, the task semantics change, or the safety controller triggers an alarm, the lease enters the revocation state.

[0121] After a lease ends or is revoked, the audit update module archives the lease event.

[0122] Step S6: Execute the action segment within the validity period of the security lease.

[0123] The execution control module executes candidate action segments when the security lease is valid, and during the execution process, it reports the action progress, status deviation, and security events to the security lease management module.

[0124] The execution control module must not execute the corresponding action segment if the lease expires, is revoked, or is not acquired.

[0125] Step S7: Trigger the downgrade policy when the lease expires or is revoked.

[0126] When a security lease expires or is revoked, the execution control module stops executing the original action segment and triggers a security stop, rollback action, low-speed conservative control, on-device re-inference, edge re-inference, cloud re-planning, or manual takeover based on the cause of the failure.

[0127] Among them, the hybrid critical level safety controls are as follows:

[0128] The hybrid critical-level safety control module runs independently of the VLA inference module, performing C0-level hard real-time safety tasks, including emergency stop, speed limit, torque limit, restricted area boundary, collision constraint, and safe shutdown.

[0129] Even if the edge network is interrupted, the cloud is unreachable, or VLA inference times out, the hybrid critical security control module can still continue to perform security controls.

[0130] In the event of a network outage, the system automatically enters offline security mode. In offline security mode, cloud-based inference is disabled, edge inference is only used when the local network is available, the lightweight ontology model and conservative skill library continue to run, and C0-level security controls maintain the highest priority. Once the network is restored, the system synchronizes scheduling events and security events based on audit logs.

[0131] The audit and updates are as follows:

[0132] The audit update module records scheduling, review, lease, and execution events during the operation.

[0133] The recorded content includes the input task, scheduling descriptor, risk-latency joint score, selected inference location, selected model specification, candidate action summary, semantic security review results, physical security review results, lease content, lease failure reason, degradation strategy, manual takeover information, and final task result.

[0134] Closed-loop audit data can be used to update model delay profiles and task risk profiles. For example, when a certain type of crawling task triggers lease revocation multiple times under high load conditions at the edge, the system increases the risk weight of this type of task under high load conditions and prioritizes short action fragments on the ontology side or requests manual confirmation.

[0135] In a specific application scenario, the method first receives the user instruction "hand the cup on the table to the designated user".

[0136] The status acquisition module acquires desktop images, depth point clouds, cup positions, user positions, robotic arm poses, and edge node loads.

[0137] The VLA inference module generates candidate action fragments for grabbing and delivering the cup.

[0138] The edge resource scheduling module generates a scheduling descriptor based on the candidate action fragments and determines that the current edge node has a high load, the personnel are close to each other, and the irreversibility of the grabbing closure action is high. Therefore, the risk level of the action fragment is marked as medium to high risk.

[0139] The scheduler calculates a risk-latency joint score based on the scheduling descriptor. When the score indicates that the edge standard model can complete inference within the latency budget, the system selects the edge standard model to generate short action segments; when the score indicates edge congestion or increased network jitter, the system selects the ontology lightweight model to execute short, conservative horizontal actions, such as approaching the target but not closing the gripper; when the score indicates that the person is too close or the obstacle is changing rapidly, the system suspends the delivery action and waits for the environment to recover.

[0140] The semantic safety governance module checks whether the task target, user identity, and delivery object conform to the rules. If the target object is an ordinary cup and the user has the right to receive it, the semantic layer approves the action; if the target object is identified as a container of dangerous liquid or the target user does not have the right to receive it, the semantic layer rejects the action and requests manual confirmation.

[0141] The physical safety constraint module predicts the grasping and delivery trajectory. If the predicted trajectory is not close enough to the human body, the physical layer rejects the action or reduces the allowable workspace; if the predicted trajectory meets the distance, speed, torque, and contact force constraints, the physical layer outputs the pass result and allowable boundaries.

[0142] Once both the semantic and physical layers are passed, the security lease management module generates a security lease for the action segment. The execution control module executes the action within the lease's validity period and continuously monitors state deviations and environmental changes. If, during execution, the system detects that a user suddenly reaches into the trajectory area, the lease cancellation condition is triggered, and the execution control module immediately stops the current action and enters a low-speed hold or rollback action.

[0143] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A risk-delay-safe scheduling method for embodied robots, characterized in that, include: S1. Obtain robot multimodal state, user task instructions, environmental semantic information, edge computing resource state, network state, and candidate action fragments generated by vision-language-action model inference; S2. Generate scheduling descriptors based on candidate action fragments and calculate the risk-latency joint score; S3. Based on the risk-latency joint score, dynamically determine the inference deployment location, model specifications, and action execution strategy; S4. Conduct semantic security governance review and physical security constraint review of candidate action segments; S5. When a candidate action fragment simultaneously satisfies the scheduling condition, semantic safety condition, and physical safety condition, a security lease is generated. S6. Execute the candidate action fragment during the validity of the safety lease; S7. When a security lease expires or is revoked, trigger the corresponding degradation strategy based on the reason for the expiration.

2. The method according to claim 1, characterized in that, In step S2, the process of generating the action fragment scheduling descriptor is as follows: 1) Extract the task semantics, candidate action trajectory, action fragment length, and safety constraints of candidate action fragments; 2) Based on the information extracted in step 1), associate the current sensor timestamp, model confidence, inference time, resource requirements, and network status; 3) Based on the information associated in step 2), combined with the risk level and available degradation strategies, generate an action segment scheduling descriptor containing the expected execution window.

3. The method according to claim 1, characterized in that, In S2, the calculation process of the risk-delay joint score is as follows: 1) Quantify the task risk, model confidence, and degree of irreversibility of the action; 2) Obtain sensor freshness, edge computing load, distance between people and obstacles, and network status; 3) Based on the end-to-end latency budget, the risk-latency joint score is calculated according to the quantification results in step 1) and the state parameters obtained in step 2).

4. The method according to claim 1, characterized in that, The specific process of S3 is as follows: S3.1 Based on the risk-latency joint score, select the robot body, edge, or cloud as the inference deployment location; S3.2 Based on the inference deployment location and task requirements, select a lightweight model, standard model, or high-precision model as the model specification; S3.3 Based on the risk-delay joint score and delay budget, determine the action segment length, execution frequency, and scheduling priority.

5. The method according to claim 1, characterized in that, In S4, the semantic security governance review process includes: 1) Parse and validate user commands, task semantics, scenario semantics, permission rules, and behavioral norms; 2) Based on the parsing and verification results of step 1), identify dangerous instructions, unauthorized instructions, ambiguous instructions, privacy risks, access to prohibited areas, abnormal interactive objects, and behaviors inconsistent with the task objectives; 3) Based on the recognition results of step 2), determine whether the candidate action meets the semantic safety requirements.

6. The method according to claim 1, characterized in that, The physical security constraint review process described in S4 includes: 1) Perform collision prediction, workspace constraints, speed constraints, acceleration constraints, torque constraints, contact force constraints, and human-robot distance constraints on candidate motion segments; 2) Based on the constraints in step 1), perform control barrier function constraints and action reversibility checks; 3) Based on the inspection results of step 2), output the physical security determination result.

7. The method according to claim 1, characterized in that, In S5, the process of generating a security lease includes: 1) The judgment results of the aggregation scheduling conditions, semantic security conditions, and physical security conditions are used to obtain a comprehensive judgment result; 2) Based on the comprehensive judgment results, generate authorization information including the effective time window, allowable state deviation, allowable risk level, and action boundaries; 3) Based on the authorization information, configure the execution frequency, degradation strategy and revocation conditions to form an executable security lease.

8. The method according to claim 7, characterized in that, The description also includes: S8. Based on the hybrid critical level, the robot's operation tasks are divided into C0 level tasks, C1 level tasks, C2 level tasks and C3 level tasks; For C0-level tasks, it runs independently of the vision-language-action model and performs emergency stops, collision constraints, speed limits, torque limits, and safety boundary monitoring in real time. For C1 level tasks, local state estimation, trajectory tracking, and low-level control are performed; For C2 level tasks, perform visual-language-action model inference, action fragment generation, and short-range task execution; For C3 level tasks, perform long-term task planning, cloud-based large model inference, log analysis, and strategy updates.

9. The method according to claim 1, characterized in that, Also includes: S9. When the network is interrupted or the edge is unavailable, cloud inference or edge inference is prohibited, and the robot body's lightweight model, conservative skill library and C0 level safety control are switched. S10. Record scheduling decisions, security reviews, lease issuance, lease cancellation, action interception, degraded execution, manual takeover, and near-miss events; S11. Based on the records in S10, update the task risk profile, model delay profile, scenario safety rules, and red team test cases.

10. A risk-delay safety scheduling system for embodied robots, characterized in that, include: The status acquisition module is used to acquire the robot's real-time perception information and task semantic information; The VLA inference module is used to generate candidate action fragments; The edge resource scheduling module is used to generate action fragment scheduling descriptors and calculate the risk-latency joint score; The semantic security governance module is used to perform semantic security reviews on candidate action fragments; The physical safety constraint module is used to perform physical safety reviews on candidate action segments; The security lease management module is used to generate a security lease when a candidate action fragment passes the review, and to revoke the security lease when the revocation conditions are met. The execution control module is used to execute action segments during the validity of the security lease and to trigger security degradation operations when the security lease expires or is revoked. The audit update module is used to record runtime events and update risk profiles and delay profiles; The hybrid critical-level safety control module operates independently of the VLA inference module and is used to perform emergency stops, collision constraints, speed limits, torque limits, and safety boundary monitoring.

Images