A vla robot training method and system based on representation invariance distillation

Abstract

The application discloses a VLA robot training method and system based on representation invariance distillation, and particularly relates to the field of artificial intelligence. A reference teacher branch is formed by an exponential moving average version of a student VLA policy network. The reference teacher branch receives original image-language input and outputs a reference action-condition representation. The student VLA policy network receives image-language input after semantic-preserving perturbation, and makes the action-condition representation close to the reference action-condition representation through representation distillation, so as to improve the robustness of the robot policy under visual perturbation, language instruction rewriting and combined perturbation.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a VLA robot training method and system based on representation-invariant distillation. Background Technology

[0002] Existing vision-language-action (VLA) robot strategies typically input current visual observations, natural language commands, and optional robot states into a vision-language backbone network, and the action generation head outputs a sequence of robot action control over a future period. With the development of large-scale vision-language pre-training, robot teaching datasets, and continuous action generation models, existing VLA models can already achieve high task success rates on standard robot operation benchmarks.

[0003] However, existing technologies primarily rely on imitation learning loss, motion trajectory regression loss, or generative motion supervision loss to train policy models, meaning they mainly require the model's output motion control sequence to closely approximate the expert-taught motion. This training method has the following shortcomings:

[0004] First, only the final motion control output is supervised, without explicitly constraining the intermediate motion condition representations upon which motion generation depends to remain stable under equivalent inputs. When image lighting, camera viewpoint, background texture, sensor noise, or language command descriptions change but the semantics of the robot's task remain unchanged, the internal motion condition representations of the model may drift, leading to instability in the motion generation process.

[0005] Second, ordinary visual data augmentation or instruction rewriting training usually treats perturbation samples as new independent training samples, and still mainly relies on the final action loss to indirectly shape the model representation. It lacks pairwise representation constraints between the original view and the perturbation view of the same sample, so it is difficult to directly guarantee that semantically equivalent inputs have consistent action condition representations before action generation.

[0006] Third, output consistency methods typically only constrain the similarity of the outputs of the original input and the perturbation input on a limited number of motion samples, velocity field sampling points, or final motion blocks. The constraint position is located at the back end of motion generation, which cannot guarantee the consistency of the condition representation before motion generation, and it is also difficult to guarantee the stability of the continuously generated motion head on the entire motion distribution.

[0007] Fourth, some robustness enhancement methods require additional sensors, historical observations, external detection modules, visual expert models, or additional branches during inference, which may increase computational overhead, inference latency, and system complexity during robot deployment.

[0008] Therefore, existing technologies still lack a robust training method that can directly constrain the VLA policy action condition representation during the training phase and without adding extra modules or inputs during the inference phase. Furthermore, existing feature distillation methods mostly focus on visual functional features such as visual encoding features, detection features, segmentation features, or depth features, while lacking a training scheme for distilling the original view to perturbed view representation for the hidden state of action queries between the visual language backbone network and the action generation head. Summary of the Invention

[0009] The purpose of this invention is to provide a VLA robot training method and system based on representation invariant distillation to solve the above-mentioned problems.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a VLA robot training method based on representation-invariant distillation, comprising:

[0011] Data acquisition and task sample representation:

[0012] Acquire robot teaching data, with each training sample including at least the current visual observation. Language instructions Optional robot states and expert action blocks .in, Indicates a time step. The time step length of the action block. For robots at all times The motion control vector, used to drive the movement of the robot actuator, may include pose control quantities of the end effector, joint space control quantities, gripper control quantities, or combinations thereof. Specifically, the pose control quantities of the end effector may include three-dimensional translation control quantities, three-dimensional rotation control quantities, or combinations thereof; the joint space control quantities may include joint angle, joint velocity, or joint torque control quantities; and the gripper control quantities may include gripper opening / closing amplitude or gripper opening / closing state.

[0013] Equivalent double view construction:

[0014] For each training sample, construct an original view and a perturbed view. The original view consists of the original visual observation and the original language instruction, i.e. ( The perturbation view consists of visual observations and linguistic instructions after semantic preservation transformation, i.e. ,in , , This represents the visual perturbation function. Represents language rewriting functions.

[0015] The visual perturbation function may include semantically preserved perturbations such as brightness, color, noise, blur, cropping, background texture, and camera view. The specific perturbation types and parameter ranges are described in the preferred implementation below.

[0016] Language rewriting functions can be obtained through pre-built offline synonym rewriting libraries, template replacement, manual review of rewriting, or filtering generated by language models. Language rewriting should maintain the object of operation, target location, spatial relationship, and intention of the action unchanged. For example, rewriting "put the carrot on the plate" as "place the carrot onto the dish".

[0017] In a preferred embodiment, the visual perturbation function The transformation includes at least one or more of the following semantically preserving visual transformations: brightness scaling, contrast scaling, color dithering, Gaussian noise, salt-and-pepper noise, slight blurring, slight cropping, proportional scaling, background texture replacement, small-range changes in camera viewpoint, or combinations thereof. Specifically, the brightness scaling factor can be between 0.6 and 1.4, the contrast scaling factor can be between 0.6 and 1.4, the standard deviation of Gaussian noise can be between 0.005 and 0.08 of the image pixel value range, the blur kernel size can be between 3 and 9, and the cropping ratio can be between 0.80 and 1.00 of the original image area. These parameter ranges are merely preferred ranges; those skilled in the art can adjust them according to camera resolution, robot task type, and target object size.

[0018] To ensure that visual perturbations do not alter the semantics of the robot's operational tasks, the visual perturbation function... Preferably, the following constraints should be met: the target object category should not be changed, the main visible area of ​​the target object should not be obscured, the key spatial relationships between the target object and the target area, such as left-right, front-back, and up-down, should not be changed, new target objects that conflict with the task instructions should not be introduced, and geometric transformations that would change the meaning of the robot's operating direction, such as horizontal flipping, large-angle rotation, or other geometric transformations, should not be performed.

[0019] Language rewriting function This is used to generate rewritten instructions that are semantically equivalent to the original language instructions. Language rewriting can be obtained through manual templates, synonym replacement, offline synonym rewriting libraries, or by generating and then filtering language models. The rewritten instructions should retain the operand, target object, target location, spatial relationships, action type, and negation / constraint conditions unchanged. Rewritten samples that do not meet the above semantic preservation conditions should be discarded or regenerated.

[0020] Student VLA and EMA Reference Faculty Branch:

[0021] The student VLA policy network consists of a visual-language backbone network and an action generation head. The visual-language backbone network is used to fuse visual observations, language instructions, and action query tags, while the action generation head is used to generate future action control sequences based on action condition representations.

[0022] The EMA reference teacher branch is an exponential moving average version of the student VLA branch parameters. If the student VLA branch parameter is θ, the teacher network parameter is... Then, after each or every few training updates, according to Update teacher network parameters, including The EMA momentum coefficient is preferably in the range of 0.90 to 0.9999. The EMA reference teacher branch does not backpropagate by representing distillation loss; instead, it receives the original view and outputs a reference action condition representation under the original view, which serves as the stable representation anchor point for the perturbation student branch.

[0023] Action query representation extraction:

[0024] Add to the input sequence Action query tag, The input sequence can be an integer greater than or equal to 1, preferably ranging from 1 to 128. The input sequence may include visual markers, linguistic markers, and action query markers. The visual-linguistic backbone network performs multi-layer attention fusion on the input sequence and outputs the hidden state at each marker location.

[0025] Select the hidden state corresponding to the action query marker from the hidden states of the last layer or the preset middle layer as the action condition representation, denoted as . This action condition representation is located between the visual language backbone network and the action generation head, and directly serves as the conditional input to the action generation head. For the EMA referencing the teacher branch, the teacher action condition representation is obtained. For the perturbed student branch, the student action condition representation is obtained. .

[0026] Indicates the invariant distillation module:

[0027] The invariant distillation module is used to constrain the action conditions of the perturbed student branch, which are close to the action conditions of the EMA reference teacher branch. Preferably, the cosine distance is used as the distillation loss. .in, and The first The action query tagged students and teachers with hidden status, This indicates that the gradient operation is stopped.

[0028] In other implementations, the distillation loss can also be represented by mean squared error loss, smoothed L1 loss, normalized feature distance, contrastive similarity loss, or a weighted combination of the above losses. To avoid over-compression of the representation, representation variance constraints, effective rank monitoring, lower bound constraints on feature standard deviation, or distillation weight scheduling strategies can be further introduced.

[0029] Action generation and action supervision module:

[0030] The action generation head can be a stream matching action generation head, a diffusion action generation head, a continuous action regression head, or other continuous action generation modules. Taking stream matching actions as an example, given an expert action block... Random noise and flow time Construct intermediate action states The action generation head represents the action conditions. Predicting the velocity field and with The loss of action is calculated as a monitoring target.

[0031] Preferably, the student's action condition is represented. Calculate student motion loss Optionally, the teacher action loss is calculated based on the teacher action conditional representation of the stopping gradient. This ensures that the shared action generation head maintains compatibility with the original representation. The ultimate training objective is... ,in , , Non-negative weights are preferred. =1, Take values ​​from 0 to 1. Take values ​​from 0.01 to 5.

[0032] Parameter update and inference deployment:

[0033] During training, the student VLA policy network and action generation head are updated via backpropagation; the teacher VLA policy network does not directly receive gradients, but is updated via the exponential moving average of the student network parameters. After training, only the student VLA policy network and action generation head are retained during the inference and deployment phase. Figure 1 As shown, solid arrows represent the forward data flow during training or inference, while dashed arrows represent training constraints, student parameter updates, or EMA parameter updates. Teacher action conditions are also represented. Student motion condition representation Both can input a shared action generation head; the shared action generation head generates or predicts action blocks based on the input action conditions, and forms teacher action supervision and student action supervision respectively. The distillation loss, teacher action supervision loss, and student action supervision loss together constitute the joint training objective.

[0034] During inference, the robot system receives current visual observations and verbal commands, inputs them into the student VLA policy network, extracts the action query hidden state as an action condition representation, and the action generation head generates a sequence of future action control for the robot to execute. The inference phase does not require a teacher network, visual perturbation module, language rewriting module, representation distillation module, or additional sensors, thus maintaining the same inference interface as the basic VLA policy.

[0035] The technical effects and advantages of this invention are as follows:

[0036] 1. This invention can reduce the drift of action condition representation between the original input and the perturbed input. By distilling the action query representation of the perturbed student branch onto the representation of the EMA reference teacher branch, this invention enables semantically equivalent image-language inputs to have a more stable condition representation before action generation.

[0037] 2. This invention enhances the robustness of VLA strategies against visual perturbations, language instruction rewriting, and combined perturbations. Since the action generation head directly depends on the action condition representation, stabilizing this interface before action generation helps reduce inconsistencies in consecutive action blocks.

[0038] 3. Compared with ordinary data augmentation training, this invention explicitly establishes a pairwise relationship between the original view and the perturbation view of the same training sample, and no longer treats the perturbation sample as an independent training sample. Therefore, it can learn the semantically invariant representation under perturbation more directly.

[0039] 4. Compared to methods that only apply consistency constraints to the output action or velocity field, this invention moves the constraint position forward to the action condition interface between the visual language backbone network and the action generation head, enabling the fixed action generation head to operate under more stable condition inputs.

[0040] 5. No additional computational branches are added during the inference phase. The EMA reference teacher branch, perturbation construction, language rewriting, and representation distillation are only used during the training phase. Only the student policy and action generation heads are retained during deployment, so they do not increase inference latency or system complexity.

[0041] 6. Applicable to various VLA architectures and action generation methods. This invention can be used in conjunction with Qwen-VL, OpenVLA-like visual language backbone networks, and action generation heads such as flow matching, diffusion, and continuous regression, exhibiting good modularity and portability.

[0042] 7. It helps prevent robust training from sacrificing performance on standard tasks. This invention preserves the action supervision loss while representing distillation, and avoids over-compression representation through teacher action loss, distillation weight scheduling, and effective rank monitoring. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the overall process of training and inference deployment of the VLA policy based on representation-invariant distillation according to the present invention.

[0044] Figure 2 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] This invention provides, for example Figures 1-2 The VLA robot training method based on representation-invariant distillation shown in the figure has the following specific steps:

[0047] S1. Acquire robot teaching data, with each training sample including at least the current visual observation, language commands, and expert action blocks;

[0048] S2, For each training sample, construct an original view and a perturbation view, wherein the original view contains the original visual observations and the original language instructions, and the perturbation view contains the visual observations and language instructions after semantic preservation transformation;

[0049] S3, input the original view to the reference teacher branch to obtain the reference action condition representation, wherein the reference teacher branch is an exponential moving average version of the student VLA policy network parameters;

[0050] S4, input the perturbation view into the student VLA policy network to obtain the student action condition representation;

[0051] S5, calculate the representation distillation loss between the student action condition representation and the reference action condition representation;

[0052] S6, Calculate the student action supervision loss based on the student action condition representation and the expert action block;

[0053] S7, based on the representation distillation loss and the student action supervision loss, update the parameters of the student VLA policy network; and

[0054] Based on the updated student VLA policy network parameters, the parameters of the reference teacher branch are updated using an exponential moving average.

[0055] Specifically, both the student VLA policy network and the reference teacher branch include a visual language backbone network and an action generation head;

[0056] The steps for obtaining the reference action condition representation and the student action condition representation specifically include:

[0057] At least one action query tag is added to the sequence input to the visual language backbone network;

[0058] From the output hidden states of the visual language backbone network, the hidden state corresponding to the action query tag is selected as the action condition representation. This action condition representation is located between the visual language backbone network and the action generation head and is directly used as the condition input of the action generation head.

[0059] 3. Specifically, the distillation loss in S5 is represented by calculating the cosine distance between the student action condition representation and the reference action condition representation, specifically:

[0060] ;

[0061] In the formula, and The first The action query tagged students and teachers with hidden status, This indicates that the gradient operation is stopped.

[0062] Specifically, the distillation loss in S5 can also be obtained by mean square error, smoothing L1 loss, normalized feature distance, contrastive similarity loss, or a weighted combination of the above losses.

[0063] Specifically, the semantic preservation transformation in S2 includes:

[0064] Semantic-preserving visual transformation, i.e.:

[0065] ;

[0066] In the formula: Represents the visual perturbation function; This represents a visual observation after semantically preserving transformation; Represents original visual observation

[0067] Semantic-preserving language rewriting, that is:

[0068] ;

[0069] In the formula: Represents language rewriting functions; These are language instructions that have undergone semantically preserved transformation. Original language instructions;

[0070] The semantically preserved visual transformations include at least one of the following: brightness scaling, contrast scaling, color jittering, Gaussian noise, salt and pepper noise, blurring, cropping, proportional scaling, background texture replacement, and camera viewpoint change.

[0071] The semantically preserved language rewriting is obtained through synonym replacement, template replacement, or language model generation followed by filtering, while maintaining the operation object, target location, spatial relationship, and action intention unchanged.

[0072] Specifically, the student action supervision loss in S6 is calculated using flow matching, diffusion model, or continuous action regression.

[0073] Specifically, when updating the parameters of the student VLA policy network, a reference teacher action supervision loss is calculated based on the reference action condition representation and the expert action block, and the reference teacher action supervision loss is included in the total loss for parameter updating.

[0074] Example 1: Verification of the stability and mission success rate of the invariant distillation scheme using the present invention.

[0075] This embodiment uses a VLA robot strategy with a visual language backbone network and a flow-matching action generation head for verification. During training, the EMA reference teacher branch receives the original image-language view, while the student VLA branch receives the image-language view with semantic preservation perturbation. The action query hidden state of the student branch is constrained by representation distillation loss to approximate the action query hidden state output by the reference teacher branch under the original view, and the student VLA branch and action generation head are updated in combination with action supervision loss.

[0076] To evaluate the stability of the action condition representation, N robot operation samples were selected from the test data, and original views and semantically preserved perturbation views were constructed for each sample. For each sample, K hidden states corresponding to action query tags were extracted. The average action condition representation distance was calculated. Defined as: For all test samples from i=1 to N and all action query tags from k=1 to K, calculate the original view hidden state respectively. With the disturbance view hidden state Cosine distance between And average the N×K cosine distances obtained. The smaller the value, the more stable the motion condition representation is between the original view and the perturbed view.

[0077] Under the current verification conditions, the average motion condition representation distance D_rep after training using the scheme of this invention is 0.0003. On the standard LIBERO robot operation benchmark, each task was executed 50 times independently. The success rates for the four task types—LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-Long—were 93.6%, 100.0%, 99.4%, and 93.6%, respectively, with an average task success rate of 96.7%. These results demonstrate that the scheme of this invention can improve the stability of motion condition representations between the original view and the perturbed view while maintaining the ability to perform tasks in standard scenes.

[0078] Example 2: Validation of the stability and task success rate of traditional augmentation training schemes

[0079] This embodiment serves as a comparative implementation, employing the same robot teaching data, visual-language backbone network, action generation head, number of training rounds, visual perturbation function, and language rewriting function as Embodiment 1. Unlike Embodiment 1, this embodiment does not include an EMA reference teacher branch, does not calculate the action query representation distillation loss, and instead directly inputs the perturbed image-language view into the student VLA policy network, updating the student VLA policy network and action generation head solely through action supervision loss. This training method corresponds to a standard visual data augmentation and language instruction rewriting training scheme.

[0080] The tests were conducted using the same test samples, perturbation construction methods, and evaluation metrics as in Example 1. Under the current validation conditions, the average action condition representation distance of the traditional augmentation training scheme... The value is 0.0310. On the standard LIBERO robot operation benchmark, each task was performed 50 times independently. The success rates of the four task types LIBERO-Spatial, LIBERO-Object, LIBERO-Goal and LIBERO-Long were 93.8%, 99.8%, 97.4% and 93.2% respectively, with an average task success rate of 96.1%.

[0081] Compared to Example 2, the average action condition representation distance in Example 1 decreased from 0.0310 to 0.0003, a reduction of 0.0307; the average task success rate increased from 96.1% to 96.7%, an increase of 0.6 percentage points. These results demonstrate that, while maintaining consistency in model structure, training data, perturbation type, and action supervision method, introducing an EMA reference teacher branch and performing representation distillation on the action query hidden state can further improve the stability of action condition representations under semantic preservation perturbations and maintain a high robot task execution success rate.

[0082] Example 3: A robust VLA training method based on stream-matched action heads

[0083] This embodiment uses a robot strategy with a visual language backbone network and a flow-matching action head as an example. The training data consists of robot teaching trajectories, each trajectory including image observations, language commands, robot states, and expert action blocks.

[0084] 1. Initialize the student VLA policy network and action generation head, and copy the student VLA policy network parameters to initialize the EMA teacher VLA policy network.

[0085] 2. Read a batch of samples from the training data to obtain visual observations. Language instructions Robot status and expert action blocks .

[0086] 3. Construct the original view from the original image and the original language. The image is subjected to semantic perturbation while maintaining visual perturbation, and the instructions are subjected to semantic rewriting while maintaining semantic perturbation, resulting in a perturbed view. .

[0087] 4. Append K action query tags after the language input in the original view and the perturbation view, and input the EMA teacher VLA and student VLA respectively.

[0088] 5. Select the hidden states corresponding to the action query markers from the hidden states of the teacher's VLA and the student's VLA, and use them as... and .

[0089] 6. According to With stopping gradient The calculation represents the distillation loss, making the student action condition representation under the perturbed input approximate the teacher action condition representation under the original input.

[0090] 7. Sampling noise and flow time ,structure The student action loss is calculated by inputting h_S into the shared action head and matching the action head; alternatively, the teacher action loss is calculated by inputting h_T into the shared action head.

[0091] 8. Based on total losses Update student VLA and motion generation head parameters.

[0092] 9. Update the teacher VLA parameters according to the EMA rules, and repeat the above steps until the training is completed.

[0093] In this embodiment, Can be 4, 8, 16, 32 or other positive integers; Action block length EMA momentum can be taken from 1 to 100. The distillation weight can be gradually increased from 0.99 to 0.9999. The parameters can be selected based on the robustness of the validation set, representation distance, and effective rank. The above parameters are merely preferred embodiments, and those skilled in the art can adjust them according to the model size, robot motion dimensions, and amount of training data.

[0094] Example 4: Training method for combining visual-language perturbations

[0095] In another embodiment, the student branch receives both visual perturbations and paraphrased language simultaneously to cover combined perturbations common in real-world robot deployments. For example, an image with altered camera viewpoint or lighting is combined with paraphrased instructions and input into the student VLA, while the teacher VLA still receives the original image and instructions. This implementation enhances the model's robustness when both visual appearance changes and verbal expression changes occur simultaneously.

[0096] Example 5: Alternative Action Generator and Alternative Distillation Position

[0097] In other embodiments, the action generation head can be replaced by a diffusion action head, a conditional action regression head, a Transformer action decoder, or other continuous action generation modules. The original view reference representation anchoring distillation method of the present invention can be used as long as the action generation module receives the action conditional representation generated by the visual language backbone network as input.

[0098] In other embodiments, in addition to the action query representation, auxiliary consistency loss can also be applied to global multimodal features, visual features, linguistic features, or the action output velocity field. However, the preferred implementation is to perform distillation on the hidden state of the action query, because this position integrates both visual and linguistic information and serves as a direct conditional input to the action generation head.

[0099] Example 6: Deployment Method

[0100] After training, the student VLA policy network and action generation head are deployed to the robot control system. In each control cycle, the robot reads the current camera image, language task instructions, and optional robot states. The student VLA policy network then obtains the hidden action query states, and the action generation head generates the future H-step action control sequence. After executing one or more actions in the action block, the robot enters the next control cycle. During the inference phase, the EMA teacher network is not invoked, and no language rewriting or visual perturbation construction is performed.

[0101] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A VLA robot training method based on representation-invariant distillation, characterized in that: The specific steps are as follows: S1. Acquire robot teaching data, with each training sample including at least the current visual observation, language commands, and expert action blocks; S2, For each training sample, construct an original view and a perturbation view, wherein the original view contains the original visual observations and the original language instructions, and the perturbation view contains the visual observations and language instructions after semantic preservation transformation; S3, input the original view to the reference teacher branch to obtain the reference action condition representation, wherein the reference teacher branch is an exponential moving average version of the student VLA policy network parameters; S4, input the perturbation view into the student VLA policy network to obtain the student action condition representation; S5, calculate the representation distillation loss between the student action condition representation and the reference action condition representation; S6, Calculate the student action supervision loss based on the student action condition representation and the expert action block; S7, based on the representation distillation loss and the student action supervision loss, update the parameters of the student VLA policy network; and: Based on the updated student VLA policy network parameters, the parameters of the reference teacher branch are updated using an exponential moving average.

2. The VLA robot training method based on representation-invariant distillation according to claim 1, characterized in that: Both the student VLA policy network and the reference teacher branch include a visual language backbone network and an action generation head. The steps for obtaining the reference action condition representation and the student action condition representation specifically include: At least one action query tag is added to the sequence input to the visual language backbone network; From the output hidden states of the visual language backbone network, the hidden state corresponding to the action query tag is selected as the action condition representation. This action condition representation is located between the visual language backbone network and the action generation head and is directly used as the condition input of the action generation head.

3. The VLA robot training method based on representation-invariant distillation according to claim 1, characterized in that: The distillation loss in S5 is represented by calculating the cosine distance between the student action condition representation and the reference action condition representation, specifically: ； In the formula, and The first The action query tagged students and teachers with hidden status, This indicates that the gradient operation is stopped.

4. The VLA robot training method based on representation-invariant distillation according to claim 1, characterized in that: The distillation loss in S5 can also be obtained by mean square error, smoothed L1 loss, normalized feature distance, comparative similarity loss, or a weighted combination of the above losses.

5. The VLA robot training method based on representation-invariant distillation according to claim 1, characterized in that: The semantic preservation transformation in S2 includes: Semantic-preserving visual transformation, i.e.: ； In the formula: Represents the visual perturbation function; This represents a visual observation after semantically preserving transformation; Represents the original visual observation; Semantic-preserving language rewriting, that is: ； In the formula: Represents language rewriting functions; These are language instructions that have undergone semantically preserved transformation. Original language instructions; The semantically preserved visual transformations include at least one of the following: brightness scaling, contrast scaling, color jittering, Gaussian noise, salt and pepper noise, blurring, cropping, proportional scaling, background texture replacement, and camera viewpoint change. The semantically preserved language rewriting is obtained through synonym replacement, template replacement, or language model generation followed by filtering, while maintaining the operation object, target location, spatial relationship, and action intention unchanged.

6. The VLA robot training method based on representation-invariant distillation according to claim 1, characterized in that: The student action supervision loss in S6 is calculated using flow matching, diffusion model, or continuous action regression.

7. The VLA robot training method based on representation-invariant distillation according to claim 1, characterized in that: When updating the parameters of the student VLA policy network, a reference teacher action supervision loss is also calculated based on the reference action condition representation and the expert action block, and the reference teacher action supervision loss is included in the total loss for parameter update.

8. A system for training a VLA robot based on representation-invariant distillation as described in claim 1, characterized in that: The system includes: The data acquisition module is used to acquire robot teaching data; The view construction module is used to construct the original view and the perturbed view for each training sample; The teacher branch module is used to receive the original view and output a reference action condition representation, the parameters of which are an exponential moving average version of the student VLA policy network parameters; The student branch module is used to receive the perturbation view and output the student action condition representation; The loss calculation module is used to calculate the representation distillation loss and the student action supervision loss; The parameter update module is used to update the parameters of the student branch module based on the representation distillation loss and the student action supervision loss, and to update the parameters of the teacher branch module based on the updated parameters.

Images