Evolutionary grabbing generation method for robot dexterous hand

Abstract

According to the technical scheme, the evolutionary grabbing generation method for the robot dexterous hand is characterized by comprising the following steps that hand posture preference optimization is introduced, grabbing adaptability is modeled into a posterior probability optimization problem, and the problem is solved; the physical perception consistency model comprises physical perception distillation used for training and physical perception sampling used for reasoning, and in the physical perception consistency model, a pre-trained diffusion model is adopted, so that the physical perception consistency model can be converged to better grabbing distribution according to positive and negative feedback iteration; and distilling the pre-trained diffusion model into a lightweight and few-step sampling model. According to the method, continuous optimization of the grabbing performance is achieved through efficient preference alignment, and the method is suitable for object grabbing tasks in simulation environments and real scenes.

Description

Technical Field

[0001] The present invention relates to an evolutionary grasping generation method for a robotic dexterous hand, belonging to the field of robotics technology. Background Art

[0002] In the grasping tasks of dexterous robotic hands in complex environments, the generalization ability is often limited due to insufficient diversity of training data. Scenarios in the real world have infinite diversity, and it is impossible to cover all possible variations through a limited dataset. In addition, existing grasping methods are mainly divided into optimization-based methods and learning-based methods. Optimization-based methods achieve a force-closure state by adjusting the hand posture, but face the problem of low computational efficiency; learning-based methods directly map from the input object to the grasping posture, but are prone to the problem of mode collapse, which limits the output diversity. At the same time, although existing diffusion model-based methods can generate diverse grasping postures, they face challenges in terms of efficiency, requiring multiple sampling steps and physical simulations, and it is difficult to achieve real-time adaptability. Summary of the Invention

[0003] The object of the present invention is to provide a grasping generation method that can learn through experience and continuously evolve according to preferences to improve the grasping performance of robots in complex environments.

[0004] To achieve the above object, the technical solution of the present invention discloses an evolutionary grasping generation method for a robotic dexterous hand, which is characterized by including the following steps:

[0005] Introduce hand posture preference optimization, model grasping adaptability as a posterior probability optimization problem, so that the physical perception consistency model can iteratively converge to a better grasping distribution according to positive and negative feedback. Among them, the physical perception consistency model includes physical perception distillation for training and physical perception sampling for inference. In the physical perception consistency model, a pre-trained diffusion model is adopted, and the pre-trained diffusion model is distilled into a lightweight, few-step sampling model.

[0006] Preferably, the physical perception consistency model introduces surface tension constraints, external penetration repulsive force constraints, and self-penetration repulsive force constraints to avoid unrealistic interactions between fingers and objects and between fingers, so as to ensure the stability, authenticity, and feasibility of the generated grasping postures.

[0007] The method disclosed by the present invention realizes continuous optimization of grasping performance through efficient preference alignment, and is applicable to object grasping tasks in simulation environments and real scenarios. Compared with the existing technical solutions, it has the following beneficial effects:

[0008] 1. Evolutionary Grasp Generation: Continuously optimize the grasping strategy through HPO, enabling the robot to learn from experience and adapt to changes in complex environments, thereby improving grasping performance;

[0009] 2. Efficient Preference Alignment: Model grasping adaptability as a posterior probability optimization problem, enabling the model to iteratively converge to a better grasping distribution based on preference signals, improving the consistency between grasping postures and human preferences;

[0010] 3. Fast Inference and Fine-tuning: PCM significantly improves the inference speed and preference fine-tuning efficiency by reducing the sampling steps and optimizing the number of iterations, enabling real-time grasp generation and reducing the computational overhead;

[0011] 4. Physical Rationality: Ensure that the generated grasping postures are geometrically and physically reasonable through physical perception distillation and sampling mechanisms, avoiding the generation of unrealistic grasping postures;

[0012] 5. Generalization Ability: Experimental results on multiple benchmark datasets show that the present invention has good generalization ability and can achieve excellent performance in both simulated environments and real scenarios;

[0013] 6. Real-time Performance: Even in the case of few-step sampling, the present invention can achieve real-time grasp generation, meeting the requirements for real-time performance in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a flowchart of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0015] The following further elaborates the present invention in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the appended claims of this application.

[0016] An evolutionary grasp generation method for a robotic dexterous hand disclosed by the present invention adopts an evolutionary grasp generation framework called "EvolvingGrasp" to continuously optimize the grasping performance through efficient preference alignment.

[0017] The core of the present invention lies in the introduction of Handpose-wise Preference Optimization (HPO), which models grasping adaptability as a posterior probability optimization problem, enabling the model to iteratively converge to a better grasping distribution based on positive and negative feedback. To further improve efficiency, the present invention also proposes a Physics-Aware Consistency Model (PCM), including physics-aware distillation (for training) and physics-aware sampling (for inference). This model achieves fast inference and efficient preference fine-tuning by pre-training a diffusion model and distilling it into a lightweight, few-step sampling model, significantly reducing the required sampling steps and optimization iterations. In addition, PCM introduces three physical constraints to ensure the stability, authenticity, and feasibility of the generated grasping poses, including surface tension constraint, external penetration repulsive force constraint, and self-penetration repulsive force constraint, thus avoiding unrealistic interactions between fingers and objects as well as between fingers.

[0018] The present invention verifies its effectiveness through multiple embodiments. The experiments were conducted on four benchmark datasets, including DexGraspNet, MultiDex, RealDex, and DexGRAB. The results show that EvolvingGrasp achieves state-of-the-art performance in both grasping success rate and sampling efficiency. Specifically, compared with existing generative methods, the present invention significantly reduces the computational time while generating high-quality, physically reasonable grasping poses, achieving a 30-fold speed improvement. In addition, by deploying the model on a real-world Shadow Hand robot, its ability to continuously optimize grasping performance through efficient preference fine-tuning in a real-world scenario is verified. For example, when the initial grasping attempt fails, the model can successfully achieve grasping through several preference fine-tuning, and the generated grasping poses can better conform to human preferences.

Claims

1. An evolutionary grasping generation method for a robot dexterous hand, characterized in that: The following steps are involved: Hand posture preference optimization is introduced, and grasp adaptability is modeled as a posterior probability optimization problem, so that the physical-perceptual consistency model can converge to a better grasp distribution according to positive and negative feedback iterations. The physical-perceptual consistency model includes physical-perceptual distillation for training and physical-perceptual sampling for reasoning. In the physical-perceptual consistency model, a pre-trained diffusion model is adopted, and the pre-trained diffusion model is distilled into a lightweight, few-step sampling model.

2. The evolutionary grasping generation method for a robot dexterous hand according to claim 1, characterized in that: The physical perception consistency model introduces surface tension constraints, external penetration repulsion constraints and self-penetration repulsion constraints to avoid unrealistic interactions between fingers and objects and between fingers, so as to ensure the stability, authenticity and feasibility of the generated grasping posture.

Images