Lightning Grasp: High Performance Procedural Grasp Synthesis with Contact Fields

TL;DR

Lightning Grasp introduces a high-performance, procedural approach to dexterous grasp synthesis by decoupling geometric computation from search via the Contact Field, enabling real-time generation of 1,000 to 10,000 diverse grasps on an $A100$ GPU. The method computes feasible contact domains on object surfaces through a BVH-accelerated Contact Field and then uses blockwise zeroth-order optimization and IK-based kinematics to realize stable grasps, avoiding manual energy-tuning and specialized templates. The approach demonstrates substantial speedups, broad object and hand compatibility, and robustness to irregular geometries, with explicit memory and implementation strategies to scale on modern GPUs. The work also discusses completeness, reuse of search results, and modularity, and releases code to accelerate future research in robotic manipulation. Overall, Lightning Grasp provides a practical, scalable pathway toward real-time dexterous grasp synthesis for diverse hands and objects, with potential for data-driven enhancements and hardware-aware design guidance.

Abstract

Despite years of research, real-time diverse grasp synthesis for dexterous hands remains an unsolved core challenge in robotics and computer graphics. We present Lightning Grasp, a novel high-performance procedural grasp synthesis algorithm that achieves orders-of-magnitude speedups over state-of-the-art approaches, while enabling unsupervised grasp generation for irregular, tool-like objects. The method avoids many limitations of prior approaches, such as the need for carefully tuned energy functions and sensitive initialization. This breakthrough is driven by a key insight: decoupling complex geometric computation from the search process via a simple, efficient data structure - the Contact Field. This abstraction collapses the problem complexity, enabling a procedural search at unprecedented speeds. We open-source our system to propel further innovation in robotic manipulation.

Figures (14)

• Figure 1: Lightning Grasp is a high-performance procedural (analytical) grasp synthesis algorithm. Compared to the other state-of-the-art analytical methods, it runs several orders of magnitude faster, produces a greater diversity of grasps, adapts to complex objects and high-DOF hands, and eliminates the need for the manual energy function tuning or template design. As shown in the figure, Lightning Grasp robustly handles highly irregular shapes with flexible, adaptable grasp poses within seconds. Code is available at https://github.com/zhaohengyin/lightning-grasp
• Figure 2: The Idea. Lightning Grasp follows three steps to grasp an object. (Left) We first identify the contact domains of each finger on the object surface, where each domain represents the feasible region a finger can reach. (Middle) We then search for an optimal set of contact points within these domains. (Right) Finally, the grasp is realized by positioning the fingers at the computed contact points.
• Figure 3: Contact Field and Its Interaction with Objects. A contact field is a collection of vectors in $\mathbb{R}^3\times \mathbb{S}^2$ (black arrows). This represents the potential contacts a hand can afford in space. Intersecting it with an object, we extract the contact domain on the object surface.
• Figure 4: BVH Representation of Contact Field. (Left) We first decompose the hand link meshes into many small contact patches. (Middle) For each contact patch, we generate its contact fields by sampling joint configurations randomly and gathering corresponding contact vectors. (Right) We build a BVH to represent the sampled contact field. Each leaf node contains several normal vectors in the leaf box.
• Figure 5: System Diagram. Our algorithm takes in hand model (kinematics structure and mesh) and object model (mesh) as input, and produces grasps. Our algorithm searches for object pose, contact point, and finger configurations sequentially. The output of each stage can be cached and reused for future forward search passes.

Theorems & Definitions (3)

• Definition 4.1: Contact Field (Point)
• Definition 4.2: Contact Field (Hand)
• Definition 4.3: Contact Surface Representation