Object grabbing action track generation method and device
By processing multimodal input information and using a conditional diffusion model to generate grasping trajectories that can adapt to any number of objects, the shortcomings of the VLA model in understanding quantity instructions are solved, enabling the robot to operate efficiently and naturally in multi-object grasping tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LION (SHENZHEN) ROBOT TECHNOLOGY CO LTD
- Filing Date
- 2026-04-02
- Publication Date
- 2026-05-12
AI Technical Summary
Existing VLA models cannot accurately understand and execute instructions that include quantity concepts, leading to behavioral degradation, unresponsiveness, or repetitive single-grab problems in multi-object grasping tasks.
By acquiring multimodal input information, processing natural language instructions and visual data, parsing quantity conditional parameters, and using a pre-trained conditional diffusion model to generate grasping action trajectories that can adapt to any number of objects, combined with a global injection mechanism and a cross-attention mechanism, continuous grasping trajectories are generated.
This technology enables robots to grasp any number of objects smoothly and simultaneously, improving control precision and robustness, expanding the range of tasks they can perform, and making them suitable for scenarios such as industrial production, logistics sorting, and home services.
Smart Images
Figure CN122008241A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and apparatus for generating the trajectory of an object grasping action. Background Technology
[0002] Visual-Language-Motion (VLA) models, as a cutting-edge direction in robotic general intelligence, aim to enable robots to interact with the physical world through natural language. Currently, the most advanced VLA models in the industry (such as Google's RT series and Physical Intelligence's...) Robotics models (such as UC Berkeley's OpenVLA) have demonstrated powerful capabilities, enabling them to understand complex language instructions and visual scenes, and directly generate continuous, multi-step robot motion trajectories, thus achieving end-to-end control. This approach of directly outputting high-frequency, dense motion sequences abandons the rigid paradigm of earlier methods that decomposed high-level instructions into predefined, discrete "action primitives," making robot movements smoother, more natural, and better adaptable to dynamic and unstructured environments.
[0003] However, despite these advanced VLA models' revolutionary breakthroughs in action representation and generation, they still suffer from fundamental flaws in understanding and executing a crucial class of instructions—instructions containing the concept of "quantity." For example, existing VLA models cannot accurately understand the meaning of quantifiers such as "three," "a," or "some" in instructions extremely common in industrial production and daily life, such as "grab three cookies from the box," "pick up a screwdriver," or "grab some apples." When faced with multi-object grasping tasks, existing VLA models typically exhibit the following failure modes: ① Behavioral degradation: incorrectly grasping only one object, ignoring the quantity requirement in the instruction. ② Unresponsiveness: due to the lack of corresponding multi-object grasping examples in the training data, the model cannot generate any meaningful actions. ③ Repeated single grasping: repeatedly calling the VLA model through external scripts or advanced planners, grasping one object at a time. This approach is inefficient and does not conform to the intuitive human behavior of grasping multiple objects at once; it is not true multi-object grasping.
[0004] Currently, the academic community has yet to find an effective method that can seamlessly and end-to-end integrate the high-level semantic concept of "quantity" into the continuous trajectory generation process of VLA models. Therefore, there is an urgent need for a new method and system that enables robots to truly understand quantity and directly generate complex, continuous multi-object grasping trajectories adaptable to any number of objects. Summary of the Invention
[0005] In view of the shortcomings of the prior art, the technical problem to be solved by the present invention is to provide a method and apparatus for generating grasping motion trajectories that enable robots to truly understand quantity and directly generate objects of any quantity.
[0006] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is to provide a method for generating the trajectory of an object grasping action, comprising the following steps: Acquire multimodal input information, which includes natural language commands and visual data of the robot's working scene; The multimodal input information is processed to obtain language features, visual features, and quantitative condition parameters; The language features, visual features, quantity condition parameters, and a random noise are input into a pre-trained conditional diffusion model to generate an object grasping motion trajectory that matches the quantity condition parameters.
[0007] Furthermore, the step of processing the multimodal input information to obtain language features, visual features, and quantity condition parameters includes the following sub-steps: The natural language instructions and visual data in the multimodal information are encoded respectively to obtain language features and visual features; The target capture quantity is parsed from the natural language instruction and used as a quantity condition parameter.
[0008] Furthermore, the method for parsing the target crawl quantity from the natural language instruction as a quantity condition parameter includes: identifying crawl quantity information in the natural language instruction; if the crawl quantity information is an exact quantity word, then directly extracting the exact quantity word as a quantity condition parameter; if the crawl quantity information is an fuzzy quantity word, then inputting the natural language instruction into a pre-trained quantity prediction model, and outputting a specific numerical value as a quantity condition parameter.
[0009] Furthermore, the step of generating an object grasping motion trajectory that matches the quantity condition parameters by inputting the language features, visual features, quantity condition parameters, and random noise into a pre-trained conditional diffusion model includes the following sub-steps: Sample random noise from a standard normal distribution; The language features, visual features, quantity condition parameters, and random noise are input into the pre-trained conditional diffusion model. The random noise is then iteratively reversed to denoise until a preset number of reverse denoising steps are completed, generating an object grasping motion trajectory that matches the quantity condition parameters.
[0010] Furthermore, the linguistic features and visual features are injected into the conditional diffusion model through a cross-attention mechanism, and the quantitative conditional parameters are injected into the conditional diffusion model through a global injection mechanism. The global injection mechanism specifically involves converting the quantity condition parameters into a condition vector, and then fusing the condition vector with the features of each layer of the conditional diffusion model during the denoising process through an addition operation.
[0011] Furthermore, the generated grasping motion trajectory adaptively adjusts its shape according to different quantity condition parameters and visual features. When the quantity condition parameters are greater than a preset threshold and the target objects indicated by the visual features are scattered in space, the grasping motion trajectory will automatically include a complex grasping strategy of first gathering and then grasping.
[0012] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is: to provide a training method for a conditional diffusion model, used to train the conditional diffusion model, comprising the following steps: Construct a training dataset, where each data point includes visual data, language instructions, expert action trajectories, and the corresponding number of targets captured; The visual data and language instructions are encoded to obtain visual features and language features; The expert action trajectory is forward diffused, and noise is added step by step to obtain a noisy trajectory with a set number of diffusion steps. The amount of noise added at each step is determined by a preset noise variance sequence. The noisy trajectory, diffusion steps, visual features, linguistic features, and target capture count are used as inputs to the conditional diffusion model to predict the noise to be added to the expert action trajectory, thus obtaining the predicted noise. A loss function is constructed based on the predicted noise and the actual added noise. The model parameters are optimized by minimizing the loss function. This process is repeated multiple times until convergence, resulting in a pre-trained conditional diffusion model.
[0013] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is to provide a robot grasping method, including the above-mentioned object grasping motion trajectory generation method, and further including the following steps: sending the generated object grasping motion trajectory to the robot's motion controller to drive the actuator to perform the grasping action.
[0014] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is: to provide an object grasping motion trajectory generation device, comprising: The information acquisition module is used to acquire multimodal input information, which includes natural language commands and visual data of the robot's working scene; The information processing module is used to process the multimodal input information to obtain language features, visual features, and quantitative condition parameters; The trajectory generation module is used to input the language features, visual features, quantity condition parameters, and a random noise into a pre-trained conditional diffusion model to generate an object grasping action trajectory that matches the quantity condition parameters.
[0015] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is: to provide a robot, including the above-mentioned object grasping motion trajectory generation device, and further including: A motion controller is used to receive the generated object grasping motion trajectory and drive the actuator based on the object grasping motion trajectory; An actuator is used to perform grasping tasks under the drive of a motion controller.
[0016] The object grasping motion trajectory generation method and apparatus of the present invention have at least the following beneficial effects: They abandon the approach of discrete action primitives and multi-level planning, and can directly generate a complete and continuous object grasping motion trajectory from natural language instructions and visual data containing quantity information in one go. This makes the robot's movements smoother and more natural, enabling it to grasp any reasonable number of objects, achieving truly end-to-end, quantity-controllable intelligent grasping; By treating the target grasping quantity as an independent, structured, and strongly conditional parameter, and continuously influencing the trajectory generation process of the diffusion model through a global injection mechanism, they ensure that the quantity instructions clearly and stably shape the macroscopic form of the entire trajectory, significantly improving controllability. The model exhibits high precision and robustness. During training, the conditional diffusion model can autonomously learn complex grasping strategies from the training data to meet different quantitative conditions. These complex strategies emerge from the model to satisfy the conditions, rather than being hard-coded by the user. Therefore, the model has good generalization ability and can adapt to reasonable combinations of quantities and scenarios not seen in the training data, generating reasonable grasping trajectories and greatly expanding the robot's task adaptability. It also solves the fundamental defects of existing advanced VLA models in executing quantitative instructions, enabling the robot to perform many real-world tasks that require operations based on specified quantities, such as industrial production, logistics sorting, and household services, demonstrating high practical value. Attached Figure Description
[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of one embodiment of the object grasping motion trajectory generation method of the present invention.
[0018] Figure 2 for Figure 1Flowchart of step S120.
[0019] Figure 3 This is a flowchart for training the conditional diffusion model.
[0020] Figure 4 for Figure 1 Flowchart of step S130.
[0021] Figure 5 This is a flowchart of one embodiment of the robot grasping method of the present invention.
[0022] Figure 6 This is a structural block diagram of one embodiment of the object grasping motion trajectory in this invention.
[0023] Figure 7 This is a structural block diagram of one embodiment of the robot of the present invention. Detailed Implementation The invention will now be further described with reference to the accompanying drawings.
[0024] Please see Figure 1 This is a flowchart of an embodiment of the object grasping motion trajectory generation method of the present invention. This embodiment includes the following steps: S110. Acquire multimodal input information. Specifically, the multimodal input information includes natural language commands and visual data of the robot's working scene. Natural language commands and visual data of the robot's working scene can be acquired in various ways. For example, natural language commands, such as "Please help me pick up 3 blocks from the table," can be received through a microphone or text input box, while real-time images of the workspace can be captured through the robot's camera.
[0025] S120. The multimodal input information is processed to obtain language features, visual features, and quantity condition parameters.
[0026] Please see Figure 2 This step S120 includes the following sub-steps: S121. The natural language instructions and visual data in the multimodal information are encoded respectively to obtain language features and visual features. Specifically, the natural language instructions and visual data are input into a VLA model, which is typically based on a combination of ViT and LLaMA architectures. ViT is a visual Transformer, specifically designed for image processing. LLaMA is a large language model, specifically designed for text processing. This combined model is responsible for receiving natural language instructions and visual data, and extracting language features and visual features from them.
[0027] S122. Parse the target quantity from the natural language instruction to obtain the quantity condition parameter. Specifically, identify the quantity information in the natural language instruction. If the quantity information is a precise quantity word, extract it directly as the quantity condition parameter. If the quantity information is a vague quantity word, input the natural language instruction into a pre-trained quantity prediction model, and output the corresponding specific value as the quantity condition parameter. The quantity prediction model can be implemented based on a small Transformer or MLP. The input data when training the quantity prediction model includes the natural language instruction and the actual target quantity to be grasped. The quantity prediction model can map some vague quantity words to specific target quantities. For example, the input natural language instruction is "take some screws," and the corresponding target quantity to be grasped is 8.
[0028] S130. Generate object grasping motion trajectory. Specifically, the language features, visual features, quantity condition parameters, and random noise are input into a pre-trained conditional diffusion model to generate an object grasping motion trajectory that matches the quantity condition parameters. The conditional diffusion model is implemented based on a U-Net architecture; please refer to [link to relevant documentation]. Figure 3 The training of the conditional diffusion model includes the following steps: a. Constructing the training dataset. Each data point in the training dataset includes visual data, language instructions, expert action trajectories, and the corresponding number of objects grasped. The expert action trajectories refer to successful object grasping action sequences and are high-quality demonstration data used to train the conditional diffusion model, ensuring that the action trajectories generated by the model are effective and reliable.
[0029] b. Encode the visual data and language instructions to obtain visual features and language features.
[0030] c. Obtaining a noisy trajectory through forward diffusion. The expert action trajectory is forward diffused, with noise added incrementally to obtain a noisy trajectory with a set number of steps. The amount of noise added at each step is determined by a preset noise variance sequence. Specifically, the forward diffusion process is defined as follows:
[0031] in, This represents a preset noise variance array. The index representing the noise variance array, i.e., the diffusion step index. . Indicates the first The trajectory contaminated by noise after initial diffusion. Indicates the first The trajectory contaminated by noise after initial diffusion. Represents the identity matrix.
[0032] By recursion, we can obtain:
[0033] in, Indicates the expert's action trajectory. Indicates cumulative noise scheduling. .
[0034] In forward diffusion, when k=K, the clean expert trajectory is transformed into pure standard normal noise (noisy trajectory). This pure noise state represents the most chaotic starting point of the model during training.
[0035] d. Using the noisy trajectory, diffusion steps, visual features, language features, and target capture quantity as inputs to the conditional diffusion model to be trained, predict the noise to be added to the expert action trajectory to obtain the predicted noise.
[0036] e. Construct a loss function based on the predicted noise and the actual added noise. Optimize the model parameters by minimizing the loss function, iterating multiple times until convergence, to obtain the pre-trained conditional diffusion model. Specifically, the loss function is:
[0037] in, Indicates the diffusion step index. Indicates visual characteristics, Indicates language features, Indicates the target number of items to be crawled. Indicates the expert's action trajectory. This represents the actual noise added. It's random noise. This represents the conditional diffusion model to be trained. Indicates prediction noise, This indicates cumulative noise scheduling.
[0038] Please see Figure 4 This step S130 includes the following sub-steps: S131. Sample a random noise sequence from a standard normal distribution. This noise is structurally pre-defined to have the same dimensions as the final object grasping trajectory to be generated. For example, if the final trajectory has T time steps and the action at each time step is D-dimensional, then the initial noise is a (T, D) tensor.
[0039] S132. Reverse denoising of random noise based on a conditional diffusion model generates an object grasping motion trajectory. The linguistic features, visual features, quantity condition parameters, and the random noise are input into the pre-trained conditional diffusion model. Iterative reverse denoising of the random noise is performed until a preset number of reverse denoising steps are completed, generating an object grasping motion trajectory that matches the quantity condition parameters. A random noise is input without giving the model any prior hints about what the trajectory should be; the trajectory is generated entirely based on the linguistic features, visual features, and quantity condition parameters.
[0040] Specifically, the linguistic and visual features are injected into the conditional diffusion model through a cross-attention mechanism, providing visual and semantic context for denoising. The quantitative conditional parameters are injected into the conditional diffusion model through a global injection mechanism. This global injection mechanism involves converting the quantitative conditional parameters into conditional vectors and then fusing these vectors with the features of each layer of the conditional diffusion model during the denoising process using an addition operation. In this operation, the quantitative conditional parameters are mapped through a simple MLP to obtain conditional vectors, which are then directly added to the embedding or residual block of each time step of the conditional diffusion model. This global injection method ensures that the quantitative conditional parameters function continuously and stably throughout the trajectory generation process.
[0041] The reverse denoising process is as follows:
[0042] in, This represents the mean function parameterized by the conditional diffusion model. It is obtained through the following formula:
[0043] By from arrive Through iteration, the original clean object grasping trajectory that matches the quantity condition parameters is finally obtained.
[0044] The generated grasping motion trajectory adaptively adjusts its shape based on different quantity condition parameters and visual features. When the quantity condition parameters exceed a preset threshold and the target objects indicated by the visual features are scattered in space, the grasping motion trajectory automatically incorporates a complex grasping strategy of first converging and then grasping. Different quantity condition parameters lead to different grasping strategies. To illustrate the impact of quantity condition parameters on the motion trajectory, an example is provided: ① When the quantity condition parameter is 1, since the trajectory of grasping a single object in the training data is usually direct and efficient, the generated trajectory will tend to be a straight line or a simple curve.
[0045] ② When the quantity condition parameter is 3 and the visual features show that the objects are scattered, since the expert trajectory for grasping multiple scattered objects in the training data usually includes the action of "gathering first", in order to satisfy the strong condition of quantity, the model will autonomously generate a complex trajectory that includes two stages of "gathering-grabbing".
[0046] ③ When the quantity condition parameter is 5, the generated trajectory will be more complex and may include multiple convergence stages, ultimately achieving the capture of five objects at once.
[0047] Please see Figure 5 This is a flowchart of an embodiment of the robot grasping method of the present invention. This embodiment includes the following steps: S100. Generate the object grasping motion trajectory. The specific steps of this step are described above in the description of the method for generating the object grasping motion trajectory, and will not be repeated here.
[0048] S200. Execute the grasping action based on the object grasping motion trajectory. Specifically, the generated object grasping motion trajectory is sent to the robot's motion controller to drive the actuator to perform the grasping action.
[0049] Please see Figure 6 This is a structural block diagram of a transposed embodiment of the object grasping motion trajectory generation method of the present invention. The object grasping motion trajectory generation device of this embodiment is used to implement the object grasping motion trajectory generation method described in the above embodiment. Specifically, the object grasping motion trajectory generation device of this embodiment includes an information acquisition module 110, an information processing module 120, and a trajectory generation module 130. Wherein: The information acquisition module 110 is used to acquire multimodal input information, which includes natural language commands and visual data of the robot's working scene.
[0050] The information processing module 120 processes the multimodal input information acquired by the information acquisition module 110 to obtain language features, visual features, and quantity condition parameters. Specifically, the information processing module 120 includes a VLA model 121 and a quantity parser 122. The VLA model 121 encodes the natural language instructions and visual data in the multimodal information to obtain language features and visual features, respectively. The quantity parser 122 parses the target capture quantity from the natural language instructions as a quantity condition parameter. In this embodiment, the quantity parser 122 has a built-in quantity prediction model.
[0051] The trajectory generation module 300 is used to input the language features, visual features, quantity condition parameters, and a random noise obtained by the information processing module 200 into a pre-trained conditional diffusion model to generate an object grasping action trajectory that matches the quantity condition parameters.
[0052] Please see Figure 7 This is a structural block diagram of a robot according to an embodiment of the present invention. Specifically, the robot of this embodiment includes the above-mentioned object grasping motion trajectory generation and transposition 100, motion controller 200, and execution mechanism 300. Wherein: The object grasping motion trajectory generation transpose 100 is used to generate the object grasping motion trajectory. Referring to the above description of the object grasping motion trajectory generation transpose, it will not be repeated here.
[0053] The motion controller 200 is used to receive the object grasping motion trajectory generated by the object grasping motion trajectory generation transpose 100 and drive the actuator based on the object grasping motion trajectory.
[0054] The actuator 300 is used to perform a grasping task under the drive of the motion controller 200.
[0055] This invention abandons the approach of discrete action primitives and multi-level programming, enabling the direct generation of complete and continuous object grasping trajectories from natural language commands and visual data containing quantity information. This makes robot movements smoother and more natural, allowing the robot to grasp any reasonable number of objects, achieving truly end-to-end, quantity-controllable intelligent grasping. This invention treats the target grasping quantity as an independent, structured, strongly conditional parameter, and continuously influences the trajectory generation process of the diffusion model through a global injection mechanism. This ensures that the quantity commands clearly and stably shape the macroscopic form of the entire trajectory, significantly improving control accuracy and robustness. During training, the conditional diffusion model can autonomously learn corresponding complex grasping strategies from the training data to meet different quantity conditions. These complex strategies emerge from the model to satisfy the conditions, rather than being hard-coded manually. Therefore, the model has good generalization ability, adapting to reasonable combinations of quantities and scenarios not seen in the training data, generating reasonable grasping trajectories, and greatly expanding the robot's task adaptability. This invention solves the fundamental defects of existing advanced VLA models in executing quantified instructions, enabling robots to perform many real-world tasks that require operations in specified quantities, such as industrial production, logistics sorting, and household services, and has high practical value.
[0056] The above description merely illustrates preferred embodiments of the present invention and is quite specific and detailed; however, it should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the appended claims.
Claims
1. A method for generating the trajectory of an object grasping motion, characterized in that, Includes the following steps: Acquire multimodal input information, which includes natural language commands and visual data of the robot's working scene; The multimodal input information is processed to obtain language features, visual features, and quantitative condition parameters; The language features, visual features, quantity condition parameters, and a random noise are input into a pre-trained conditional diffusion model to generate an object grasping motion trajectory that matches the quantity condition parameters.
2. The method for generating the trajectory of object grasping action as described in claim 1, characterized in that, The step of processing the multimodal input information to obtain language features, visual features, and quantitative condition parameters includes the following sub-steps: The natural language instructions and visual data in the multimodal information are encoded respectively to obtain language features and visual features; The target capture quantity is parsed from the natural language instruction and used as a quantity condition parameter.
3. The method for generating the trajectory of object grasping action as described in claim 2, characterized in that, The method for parsing the target crawl quantity from the natural language instruction as a quantity condition parameter includes: identifying crawl quantity information in the natural language instruction; if the crawl quantity information is an exact quantity word, then directly extracting the exact quantity word as a quantity condition parameter; if the crawl quantity information is an fuzzy quantity word, then inputting the natural language instruction into a pre-trained quantity prediction model, and outputting a specific numerical value as a quantity condition parameter.
4. The method for generating the trajectory of object grasping action as described in claim 1, characterized in that, The step of generating an object grasping motion trajectory that matches the quantity condition parameters by inputting the language features, visual features, quantity condition parameters, and random noise into a pre-trained conditional diffusion model includes the following sub-steps: Sample random noise from a standard normal distribution; The language features, visual features, quantity condition parameters, and random noise are input into the pre-trained conditional diffusion model. The random noise is then iteratively reversed to denoise until a preset number of reverse denoising steps are completed, generating an object grasping motion trajectory that matches the quantity condition parameters.
5. The method for generating the trajectory of object grasping action as described in claim 1, characterized in that: The linguistic and visual features are injected into the conditional diffusion model through a cross-attention mechanism, and the quantitative conditional parameters are injected into the conditional diffusion model through a global injection mechanism. The global injection mechanism specifically involves converting the quantity condition parameters into a condition vector, and then fusing the condition vector with the features of each layer of the conditional diffusion model during the denoising process through an addition operation.
6. The method for generating the trajectory of object grasping motion as described in claim 5, characterized in that: The generated grasping motion trajectory adaptively adjusts its shape according to different quantity condition parameters and visual features. When the quantity condition parameters are greater than a preset threshold and the target objects indicated by the visual features are scattered in space, the grasping motion trajectory will automatically include a complex grasping strategy of first gathering and then grasping.
7. A method for training a conditional diffusion model, used to train the conditional diffusion model as described in any one of claims 1-6, characterized in that, Includes the following steps: Construct a training dataset, where each data point includes visual data, language instructions, expert action trajectories, and the corresponding number of targets captured; The visual data and language instructions are encoded to obtain visual features and language features; The expert action trajectory is forward diffused, and noise is added step by step to obtain a noisy trajectory with a set number of diffusion steps. The amount of noise added at each step is determined by a preset noise variance sequence. The noisy trajectory, diffusion steps, visual features, linguistic features, and target capture count are used as inputs to the conditional diffusion model to predict the noise to be added to the expert action trajectory, thus obtaining the predicted noise. A loss function is constructed based on the predicted noise and the actual added noise. The model parameters are optimized by minimizing the loss function. This process is repeated multiple times until convergence, resulting in a pre-trained conditional diffusion model.
8. A robot grasping method, characterized in that, The object grasping motion trajectory generation method as described in any one of claims 1-6 further includes the following step: sending the generated object grasping motion trajectory to the robot's motion controller to drive the actuator to perform the grasping motion.
9. A device for generating the trajectory of an object grasping motion, characterized in that, include: The information acquisition module is used to acquire multimodal input information, which includes natural language commands and visual data of the robot's working scene; The information processing module is used to process the multimodal input information to obtain language features, visual features, and quantitative condition parameters; The trajectory generation module is used to input the language features, visual features, quantity condition parameters, and a random noise into a pre-trained conditional diffusion model to generate an object grasping action trajectory that matches the quantity condition parameters.
10. A robot, characterized in that, The device for generating the trajectory of object grasping motion as described in claim 9 further includes: A motion controller is used to receive the generated object grasping motion trajectory and drive the actuator based on the object grasping motion trajectory; An actuator is used to perform grasping tasks under the drive of a motion controller.