Główne pojęcia
GraspXL is a reinforcement learning framework that can synthesize diverse grasping motions for a wide range of unseen objects while adhering to specific motion objectives such as graspable areas, heading directions, wrist rotations, and hand positions.
Streszczenie
The paper presents GraspXL, a reinforcement learning-based framework for generating grasping motions that can adhere to various motion objectives, including graspable areas, heading directions, wrist rotations, and hand positions. The key highlights are:
GraspXL can synthesize grasping motions for over 500,000 unseen objects without relying on any 3D hand-object interaction data during training. This is a significant improvement in scalability compared to existing methods.
The framework introduces a learning curriculum and an objective-driven guidance technique to enable the policy to learn stable grasping while satisfying multiple motion objectives. This helps the policy avoid getting stuck in local optima caused by the conflicting objectives.
GraspXL is general enough to be deployed on reconstructed or generated objects, as well as different dexterous hand platforms, such as Shadow, Allegro, and Faive, demonstrating its broad applicability.
Extensive experiments show that GraspXL outperforms existing methods in terms of success rate, objective error, and contact ratio on both PartNet and ShapeNet datasets. It also maintains strong performance on the large-scale Objaverse dataset, with an average success rate of 82.2%.
Ablation studies highlight the importance of the learning curriculum, objective-driven guidance, and joint distance features in achieving the superior performance of GraspXL.
Statystyki
The PartNet test set contains 48 unseen objects, while the ShapeNet test set contains 3,993 unseen objects.
The Objaverse dataset used for large-scale evaluation contains over 500,000 objects.
Cytaty
"Our method does not rely on any 3D hand-object data to train but can robustly generalize to grasp a broad range of unseen objects."
"We formulate GraspXL in the reinforcement learning paradigm and leverage physics simulation."
"We introduce a learning curriculum to decompose the learning process to objective learning and grasp learning."