Abstract
A dexterous hand capable of grasping any object is essential for the development of general-purpose embodied intelligent robots. However, due to the high degree of freedom in dexterous hands and the vast diversity of objects, generating high-quality, usable grasping poses in a robust manner is a significant challenge. In this paper, we introduce DexGrasp Anything, a method that effectively integrates physical constraints into both the training and sampling phases of a diffusion-based generative model, achieving state-of-the-art performance across nearly all open datasets. Additionally, we present a new dexterous grasping dataset containing over 3.4 million diverse grasping poses for more than 15k different objects, demonstrating its potential to advance universal dexterous grasping.
Introduction
While simple parallel-jaw grippers are common in industrial settings, achieving universal robotic dexterity—using multi-fingered hands to manipulate arbitrary objects—remains a primary goal for general-purpose robotics. This task is inherently difficult due to the high degrees of freedom (DoF) in human-like hands and the infinite variety of object geometries. Existing approaches often struggle with the massive search space or produce "ghost" grasps that appear visually plausible but fail physically due to object penetration or lack of contact. DexGrasp Anything addresses these challenges by embedding physical awareness directly into a generative diffusion framework and leveraging a massive-scale dataset.
Methodology
The core of the framework is a physics-aware diffusion model that treats grasp generation as a denoising process guided by the laws of physics.
Physics-Aware Diffusion
Unlike standard diffusion models that only learn to match a data distribution, this model incorporates three specific physical forces into its loss function (Physics-Aware Training) and its inference process (Physics-Guided Sampler):
- Surface Pulling Force: Directs the hand's inner surface toward the object to ensure stable contact.
- External-Penetration Repulsion: Prevents finger segments from clipping through the object geometry.
- Self-Penetration Repulsion: Prevents internal collisions between the hand's own fingers.
LLM-Enhanced Representations
To provide semantic context beyond raw geometry, the system uses a Large Language Model (LLM) to generate descriptive features of the target object. By combining these semantic descriptions with 3D point cloud data, the model gains a functional understanding of the object (e.g., identifying the neck of a detergent bottle), which aids in proposing more intuitive and stable grasp poses.
The DexGrasp Anything (DGA) Dataset
A significant contribution of this work is the DGA Dataset, the largest and most diverse of its kind. It contains over 3.4 million grasping poses for more than 15,000 unique objects. Every pose in the dataset has been strictly validated in the IsaacGym physics simulator to ensure stability and minimal penetration. The authors utilized a "model-in-the-loop" scaling approach, where their own trained model was used to self-correct and filter new data, enabling exponential scaling of high-quality training samples.
Results and Performance
DexGrasp Anything consistently outperforms existing state-of-the-art (SOTA) models across five major benchmarks: DexGraspNet, UniDexGrasp, MultiDex, RealDex, and DexGRAB.
Key performance highlights include:
- Success Rates: On the DexGraspNet benchmark, the model achieved a 57.5% success rate (Suc.6), significantly higher than previous models like UGG (46.9%) and SceneDiffuser (26.6%).
- Physical Realism: The penetration (Pen.) metric was drastically reduced, indicating that the physics-aware training successfully taught the model to respect physical boundaries.
- Generalization: The model demonstrated a robust ability to grasp "unseen" objects from the Objaverse dataset, proving it has learned generalizable grasping principles rather than simple pose memorization.
- Real-World Deployment: The model was successfully validated on a real ShadowHand robot, showing smooth pre-grasping and lifting sequences for complex objects.
Conclusion and Future Work
DexGrasp Anything represents a major step toward universal robotic dexterity by treating physics as a fundamental prior in generative modeling rather than a post-processing step. While the model excels at generating static grasp poses, future work will focus on integrating complex motion planning for dynamic trajectories and incorporating real-time tactile feedback to enhance reliability during physical execution. The release of the DGA dataset and the physics-aware framework provides a powerful foundation for the next generation of embodied intelligent robots.