Towards a Multi-Embodied Grasping Agent

2025-10-31

AI Summary

Multi-arm grasping suffers from poor generalization across diverse robotic manipulator configurations and heavy reliance on large-scale annotated datasets. Method: This paper proposes a geometry-aware equivariant grasp generation framework. We introduce, for the first time, a batched equivariant neural network architecture implemented in JAX, explicitly encoding manipulator kinematics. The model derives kinematic constraints solely from scene point clouds and end-effector geometric structure, enabling parallel batch processing of scenes, manipulators, and grasp poses. By innovatively integrating flow-based equivariant networks with geometric deep learning, it achieves degree-of-freedom-agnostic generalization. Results: Evaluated on a large-scale dataset comprising 25,000 scenes and 20 million grasp samples, our method significantly improves inference speed and grasp success rate. Crucially, it demonstrates strong cross-configuration generalization—achieving high performance on previously unseen manipulator geometries without retraining.

Technology Category

• Intelligent Robots: Manipulation
• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Machine Learning: Learning with Manifolds

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Multi-embodiment grasping focuses on developing approaches that exhibit generalist behavior across diverse gripper designs. Existing methods often learn the kinematic structure of the robot implicitly and face challenges due to the difficulty of sourcing the required large-scale data. In this work, we present a data-efficient, flow-based, equivariant grasp synthesis architecture that can handle different gripper types with variable degrees of freedom and successfully exploit the underlying kinematic model, deducing all necessary information solely from the gripper and scene geometry. Unlike previous equivariant grasping methods, we translated all modules from the ground up to JAX and provide a model with batching capabilities over scenes, grippers, and grasps, resulting in smoother learning, improved performance and faster inference time. Our dataset encompasses grippers ranging from humanoid hands to parallel yaw grippers and includes 25,000 scenes and 20 million grasps.

Problem

• Developing generalist grasping approaches for diverse gripper designs
• Addressing data scarcity challenges in multi-embodiment robotic grasping
• Creating efficient grasp synthesis for variable gripper kinematics

Innovation

• Flow-based equivariant grasp synthesis architecture
• Handles diverse grippers with variable degrees
• JAX implementation enables batching and faster inference