Toward a Multi-Embodied Grasping Agent

Multi-embodiment grasping aims to develop approaches that exhibit generalist behavior across diverse gripper designs. Existing methods often learn the gripper kinematic structure 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 handles different gripper types with variable degrees of freedom and exploits the underlying kinematic model, deducing all necessary information solely from gripper and scene geometry. Unlike previous equivariant grasping methods, we implement all modules in JAX and provide batching capabilities over scenes, grippers, and grasps, resulting in smoother learning, improved performance, and faster inference. Our dataset encompasses grippers ranging from humanoid hands to parallel-jaw designs, including 25,000 scenes and 20 million grasps.