Robot AI control system fusing physical constraints

Abstract

The invention discloses a robot AI control system fused with physical constraints, and the system comprises the following steps: S1, employing a differentiable constraint loss layer which comprises kinematics constraint loss, dynamics constraint loss, collision-free constraint loss and contact constraint loss, S2, calculating a total loss function as the weighted sum of task loss and each constraint loss, and S3, estimating the gradient of the non-micro constraint by adopting a finite difference method. According to the method, a differential constraint loss layer is adopted in a training stage by adopting a double-level constraint fusion mechanism, a loss item based on physical constraint is introduced, gradient updating is directly guided, a constraint layer based on classifier guidance is matched, and for a controller based on a diffusion model and other iteration generation models, in the sampling process of denoising, the noise is reduced, and the noise is reduced. And a constraint function gradient is utilized to guide the generation direction, so that the device has high safety guarantee, the output quality is improved, the generalization ability is enhanced, and the modularization and flexibility are improved.

Description

Technical Field

[0001] This invention relates to the fields of robot control and artificial intelligence technology, and more specifically, to a robot AI control system that integrates physical constraints. Background Technology

[0002] Deep learning-based artificial intelligence technology has brought revolutionary progress to robot control, enabling robots to handle perception and decision-making tasks in complex environments. However, most current data-driven robot control models (such as imitation learning and reinforcement learning models) share a common flaw: their learning objectives are usually focused only on task completion, lacking explicit modeling and adherence to the underlying physical constraints.

[0003] These physical constraints include:

[0004] Kinematic constraints: such as joint angle and velocity limits, to avoid falling into unusual configurations;

[0005] Dynamic constraints: such as joint torque limits and actuator output force limits;

[0006] Contact constraints: such as the requirement that the force be closed, there be no slippage, or there be a friction cone when manipulating an object;

[0007] Environmental constraints: such as avoiding collisions with obstacles.

[0008] Ignoring these constraints can lead to a series of serious problems:

[0009] 1. Output command not feasible: The model may output actions that exceed the joint limit or motor torque range, resulting in control failure.

[0010] 2. Hardware damage risk: Abnormal movements with high speed or high torque may cause physical damage to the robot body and external objects.

[0011] 3. Behavior violates physical laws: The model may plan to penetrate obstacles or violate contact mechanics, causing the simulation to differ from the real world (Sim2RealGap).

[0012] 4. Unstable training: In reinforcement learning, invalid actions will reduce sample efficiency and prolong training time.

[0013] Existing solutions mostly process the data in the backend, such as post-correcting the model's output actions through filters and constraint optimizers. However, this approach is an "open-loop" compensation, and the correction process may deviate from the model's original intent, failing to guide the model to generate more reasonable strategies during the training phase. Therefore, there is an urgent need in this field for a "closed-loop" framework that can internalize physical constraints into the model learning process. Summary of the Invention

[0014] To address the problems existing in the prior art, the present invention aims to provide a robot AI control system that integrates physical constraints. By adopting a two-level constraint fusion mechanism and using a differentiable constraint loss layer in the training phase, the system achieves full lifecycle management of physical constraints from offline design to online adaptation, thus solving the problems in the background technology.

[0015] To achieve the above objectives, the present invention adopts the following technical solution.

[0016] A robot AI control system that integrates physical constraints includes the following steps:

[0017] S1. A differentiable constraint loss layer is adopted, including kinematic constraint loss, dynamic constraint loss, collision-free constraint loss and contact constraint loss;

[0018] S2. The total loss function is calculated as the weighted sum of the task loss and the loss of each constraint.

[0019] S3. Estimate the gradient using the finite difference method for non-differentiable constraints;

[0020] S4. Employ a hierarchical processing architecture for constraint priority management.

[0021] S5, Online Constraint Dynamic Correction Layer, adopts real-time sensor fusion, multi-data source alignment, and Kalman filtering to synchronize visual, force and proprioceptive data;

[0022] S6. Based on the synchronized data, establish an emergency constraint trigger protection mechanism.

[0023] As a further description of the above technical solution: steps S1 and S2 are used in the training phase, and steps S3, S4, S5 and S6 are used in the inference phase.

[0024] As a further description of the above technical solution: the kinematic constraint loss includes:

[0025] The formula for joint limitation loss is:

[0026] ;

[0027] in: The actual angle of the i-th joint The upper limit of the angle of the i-th joint. This is the lower limit of the angle of the i-th joint;

[0028] Singularity loss avoidance: ;

[0029] in: For the Jacobian matrix of the robot, It is a positive number.

[0030] As a further description of the above technical solution: the formula for the dynamic constraint loss is:

[0031] ;

[0032] in, The actual torque of the i-th joint. This is the upper limit of the torque of the i-th joint.

[0033] As a further description of the above technical solution: the formula for the collision-free constraint loss is:

[0034] ;

[0035] in, This represents the actual distance between the robot's links and the obstacle. This is a preset safe distance; anything less than this distance is considered dangerous.

[0036] As a further description of the above technical solution: the contact constraint loss:

[0037] ;

[0038] in: Force in the tangential direction, The coefficient of friction indicates the roughness of the contact surface. Force in the normal direction.

[0039] As a further description of the above technical solution: the guiding method in step S3 includes: in each denoising iteration, calculating the gradient of the constraint condition C(x) with respect to the current state x. ;

[0040] This gradient is used to adjust the denoising direction, causing the generated samples to move towards the region that satisfies the constraints. The expression is:

[0041] ;

[0042] in: The noise predicted by the model at time step t, This is a scaling factor related to time step t, used to balance the intensity of noise correction. The corrected noise is used to guide the generation process.

[0043] As a further description of the above technical solution: the protection mechanism in step S6 is divided into:

[0044] A yellow warning indicates a minor violation of constraints, in which QP optimization and correction are performed; an orange warning indicates a persistent violation, in which a degraded operating mode is implemented; a red warning indicates a hardware hazard, in which an emergency stop circuit is triggered.

[0045] Compared with the prior art, the advantages of this invention are:

[0046] (1) This scheme adopts a two-level constraint fusion mechanism and uses a differentiable constraint loss layer in the training stage to introduce a loss term based on physical constraints, which directly guides the gradient update. Combined with a constraint layer based on classifier guidance, for controllers based on diffusion model and other iterative generation models, the constraint function gradient guides the generation direction during the denoising sampling process, thereby achieving the advantages of high security, improved output quality, enhanced generalization ability, and improved modularity and flexibility.

[0047] (2) This scheme provides the model with strong physical priors through the constraint loss term, which guides it to find a solution that meets both the task requirements and physical laws more quickly, thus improving training efficiency. At the same time, it adds a closed loop of constraint guarantee to realize the whole link of "training-inference-execution" to cope with sudden environmental changes, thereby realizing the complete life cycle management of physical constraints from offline design to online adaptation of the device. Attached Figure Description

[0048] Figure 1 This is a schematic diagram illustrating the operating principle of the present invention. Detailed Implementation

[0049] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0050] To address the issue that most current data-driven robot control models typically focus solely on task completion as their learning objective, lacking explicit modeling and adherence to underlying physical constraints, Example 1 is proposed:

[0051] Please see Figure 1 In this embodiment, a robot AI control system that integrates physical constraints includes the following steps:

[0052] S1. A differentiable constraint loss layer is adopted, including kinematic constraint loss, dynamic constraint loss, collision-free constraint loss and contact constraint loss;

[0053] S2. The total loss function is calculated as the weighted sum of the task loss and the loss of each constraint.

[0054] S3. Estimate the gradient using the finite difference method for non-differentiable constraints;

[0055] S4. Employ a hierarchical processing architecture for constraint priority management;

[0056] S5, Online Constraint Dynamic Correction Layer, adopts real-time sensor fusion, multi-data source alignment, and Kalman filtering to synchronize visual, force and proprioceptive data;

[0057] S6. Based on the synchronized data, establish an emergency constraint trigger protection mechanism.

[0058] Steps S1 and S2 are used in the training phase, and steps S3, S4, S5 and S6 are used in the inference phase.

[0059] The kinematic constraint loss includes:

[0060] The formula for joint limitation loss is:

[0061] ;

[0062] in: The actual angle of the i-th joint The upper limit of the angle of the i-th joint. Define the lower limit of the angle of the i-th joint; calculate the angle of each joint. For portions exceeding the permissible range, if the joint angle exceeds the upper limit. or below the lower limit If the excess is summed, the loss is calculated.

[0063] Singularity loss avoidance: ;

[0064] in: For the Jacobian matrix of the robot, A positive number measures how close the robotic arm is to a singular position (i.e., a point where it loses dexterity), calculated by the Jacobian matrix. The smaller the determinant value, the closer it is to a singularity, and the greater the loss.

[0065] The formula for the dynamic constraint loss is:

[0066] ;

[0067] in, The actual torque of the i-th joint. Calculate the torque of each joint as the upper limit of the torque of the i-th joint. Exceeding the maximum allowed value The loss is calculated by summing the excess values ​​of all joints.

[0068] The formula for the collision-free constraint loss is:

[0069] ;

[0070] in, This represents the actual distance between the robot's links and the obstacle. This is a preset safe distance; anything less than this distance is considered dangerous. Specifically, it checks the distance between the robot and obstacles. If the distance Less than the safe distance Then calculate the difference between the safe distance and the actual distance, and sum the results for all dangerous situations as the loss.

[0071] The formula for the contact constraint loss is:

[0072] ;

[0073] in: Force in the tangential direction, The coefficient of friction indicates the roughness of the contact surface. The force in the normal direction specifically checks whether the contact force satisfies the friction cone constraint (i.e., tangential force). (Not too large). If the tangential force... Exceeding the coefficient of friction With normal force The product of the two is used to calculate the excess portion, and the summation is applied over all contact points as the loss.

[0074] In this embodiment, by deeply embedding physical constraints such as kinematics, dynamics, contact, and environment into the entire training and inference process of the AI ​​model, it is ensured from the source that the action commands output by the AI ​​model conform to the hardware capabilities and physical laws. This greatly reduces the risk of hardware damage and behavioral failure. The actions output by the model are inherently smooth, stable, and reliable, reducing the need for backend repair processing and improving the overall performance of the system. The model, which internalizes physical laws, can make more reasonable inferences when facing new environments and new objects, exhibiting better generalization. Different constraint modules can be flexibly selected and combined according to the specific robot and task requirements, making it widely applicable. Thus, the device has the advantages of high safety assurance, improved output quality, enhanced generalization ability, and improved modularity and flexibility.

[0075] Furthermore, to address the issues of hardware damage and control failure caused by the neglect of physical constraints in traditional data-driven models, a second embodiment is proposed:

[0076] Please see Figure 1 The guiding method in step S3 includes: calculating the gradient of constraint C(x) with respect to the current state x in each denoising iteration. .

[0077] This gradient is used to adjust the denoising direction, causing the generated samples to move towards the region that satisfies the constraints. The expression is:

[0078] ;

[0079] in: The noise predicted by the model at time step t, This is a scaling factor related to time step t, used to balance the intensity of noise correction. The corrected noise is used to guide the generation process; in general, during the denoising process, the noise direction is adjusted according to the gradient of the constraint C(x). First, the gradient of the constraint under the current state x is calculated. Then, the gradient is used to correct the noise in the model's predictions. This makes the generated samples more satisfied with the constraints.

[0080] The protection mechanism in step S6 is divided into:

[0081] A yellow warning indicates a minor violation of constraints, in which QP optimization and correction are performed; an orange warning indicates a persistent violation, in which a degraded operating mode is implemented; a red warning indicates a hardware hazard, in which an emergency stop circuit is triggered.

[0082] In this embodiment, the model is provided with strong physical priors through the constraint loss term, which guides it to find a solution that meets both task requirements and physical laws more quickly, thereby improving training efficiency. At the same time, a new closed loop of constraint protection is added to realize the entire "training-inference-execution" chain, which can cope with sudden environmental changes. Thus, the device has realized the full life cycle management of physical constraints from offline design to online adaptation.

[0083] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A robot AI control system integrating physical constraints, characterized in that: Includes the following steps: S1. A differentiable constraint loss layer is adopted, including kinematic constraint loss, dynamic constraint loss, collision-free constraint loss and contact constraint loss; S2. The total loss function is calculated as the weighted sum of the task loss and the loss of each constraint. S3. Estimate the gradient using the finite difference method for non-differentiable constraints; S4. A hierarchical processing architecture is adopted for constraint priority management. S5, Online Constraint Dynamic Correction Layer, adopts sensor fusion, multi-data source alignment, and Kalman filtering to synchronize visual, force and proprioceptive data; S6. Based on the synchronized data, establish an emergency constraint trigger protection mechanism.

2. The robot AI control system integrating physical constraints according to claim 1, characterized in that: Steps S1 and S2 are used in the training phase, and steps S3, S4, S5 and S6 are used in the inference phase.

3. The robot AI control system integrating physical constraints according to claim 1, characterized in that: The kinematic constraint loss includes: The formula for joint limitation loss is: ； in: The actual angle of the i-th joint The upper limit of the angle of the i-th joint. This is the lower limit of the angle of the i-th joint; Singularity avoidance of losses: ; in: For the Jacobian matrix of the robot, It is a positive number.

4. The robot AI control system integrating physical constraints according to claim 1, characterized in that: The formula for the dynamic constraint loss is: ； in, The actual torque of the i-th joint. This is the upper limit of the torque of the i-th joint.

5. A robot AI control system integrating physical constraints according to claim 1, characterized in that: The formula for the collision-free constraint loss is: ； in, This represents the actual distance between the robot's links and the obstacle. This is a preset safe distance; anything less than this distance is considered dangerous.

6. The robot AI control system integrating physical constraints according to claim 1, characterized in that: The contact constraint loss: ； in: Force in the tangential direction, The coefficient of friction indicates the roughness of the contact surface. Force in the normal direction.

7. A robot AI control system integrating physical constraints according to claim 1, characterized in that: The guidance method in step S3 includes: in each denoising iteration, calculating the gradient of the constraint C(x) with respect to the current state x. ; This gradient is used to adjust the denoising direction, causing the generated samples to move towards the region that satisfies the constraints. The expression is: ； in: The noise predicted by the model at time step t, This is a scaling factor related to time step t, used to balance the intensity of noise correction. The corrected noise is used to guide the generation process.

8. A robot AI control system integrating physical constraints according to claim 1, characterized in that: The protection mechanism in step S6 is divided into: A yellow warning indicates a minor violation of constraints, in which QP optimization and correction are performed; an orange warning indicates a persistent violation, in which a degraded operating mode is implemented; a red warning indicates a hardware hazard, in which an emergency stop circuit is triggered.

Images