dexterous grasp synthesis

This work addresses the disconnect between dexterous hand design and task-driven control, as well as the limited optimization dimensions in existing approaches, by proposing a unified parametric co-design framework that jointly optimizes palm structure, finger kinematics, fingertip geometry, and surface curvature. Innovatively incorporating fine-grained geometric features—such as surface curvature—into the design space, the method employs a parametric surface deformation kernel to directly model contact interactions, enabling end-to-end simulation-to-reality optimization. Integrating multi-degree-of-freedom finger modeling, task-oriented optimization algorithms, and manufacturability constraints, the approach significantly enhances grasp stability in dynamic manipulation tasks and produces dexterous hand models ready for both simulation and physical fabrication. The resulting designs are open-sourced to advance research in co-design and cross-platform policy transfer.

Existing dexterous grasp synthesis methods struggle to simultaneously achieve high success rates and diverse contact patterns across objects of varying scales. To address this challenge, this work proposes the HUGS framework, which innovatively leverages a small-scale human grasp dataset to learn a conditional prior that adaptively recommends both contact modes and initial wrist poses. Integrated with force-closure-aware optimization, HUGS unifies the generation of multimodal grasps—ranging from precision pinches to bimanual coordination. The approach overcomes limitations of traditional heuristic strategies, synthesizing 3.2 million grasp samples across 157,000 scenes covering objects from 2 to 30 cm in size. Real-world experiments demonstrate that the system autonomously selects appropriate contact modes, successfully grasping diverse objects—from small screws to large boxes—with high reliability.

Current dexterous hand design and control are largely decoupled, lacking unified evaluation metrics and co-optimization frameworks. Method: We propose the first end-to-end co-design framework that jointly optimizes hand morphology—fully parameterized across joints, fingers, and palm—and morphology-conditioned reinforcement learning control policies. Our approach integrates differentiable dynamics modeling, modular structural generative networks, and an automated simulation-to-reality transfer pipeline compatible with off-the-shelf mechanical components. Contribution/Results: We introduce a novel cross-morphology control evaluation mechanism, substantially enhancing design-space scalability. Empirical validation on in-hand rotation tasks demonstrates that the framework completes custom dexterous hand design, policy training, and physical deployment within 24 hours—achieving an efficient, closed-loop workflow from parametric search to real-world embodiment.

This work addresses the challenge in anthropomorphic dexterous hand design where morphological, actuation, and sensing parameters are highly coupled and lack systematic multi-objective optimization methods. To tackle this, the authors propose a modular-finger-based multi-parameter benchmarking framework. Through modular mechanical design, the framework enables multidimensional quantitative evaluation of key components—including joints, skeletal structure, skin compliance, and sensor placement—and establishes quantitative relationships between mechanism-level characteristics and task-level performance. Optimized finger modules are then integrated into a teleoperated full hand for task-level validation. Experimental results demonstrate that the resulting high-performance dexterous hand significantly outperforms baseline designs in tasks such as multi-object grasping and bulb screwing, confirming that finger-level co-optimization effectively enhances overall hand dexterity.

Existing functional grasping methods predict only coarse interaction regions, making it difficult to directly constrain 6D grasp poses and resulting in a disconnect between visual perception and dexterous manipulation. To bridge this gap, we propose Contact-anchored Multi-keypoint Affordance Representation (CMKA), which explicitly encodes task-driven functional contact points as anchors for 6D grasp pose estimation. Our approach introduces two key innovations: (1) a contact-guided weakly supervised keypoint learning mechanism, and (2) keypoint-guided grasp transformation (KGT), jointly leveraging weak supervision from human grasping images, fine-grained features from large vision models, geometric priors, and robot kinematic mappings to ensure hand-object spatial consistency. Evaluated on the FAH dataset, IsaacGym simulations, and real-robot experiments, CMKA significantly improves affordance localization accuracy, grasp pose consistency, and generalization across diverse tools and manipulation tasks.

This work investigates how large-scale grasping data can enhance robotic dexterity in complex tool manipulation tasks, moving beyond their conventional use for basic grasping. The authors construct a pretraining dataset comprising 355,000 grasp trajectories and introduce a hierarchical imitation learning framework: a high-level policy predicts hand subgoals, while a low-level controller executes goal-conditioned actions. By combining pretraining on this dataset with fine-tuning on downstream tasks, the approach enables efficient policy transfer. This study presents the first demonstration that large-scale grasping data can effectively pretrain contact-rich dexterous manipulation policies. The method significantly outperforms both end-to-end diffusion-based strategies and hierarchical baselines trained from scratch on the DexCraft simulation benchmark and in real-world experiments, achieving a 33.3 percentage point improvement in task success rate over DP3.

This work addresses the challenge of open-loop grasping under uncertainty in object shape and pose, where poor contact coordination often leads to slippage or failure. The authors propose a tactile feedback–based model predictive controller that enables coordinated multi-contact interaction and adaptive force modulation during both approach and grasp phases. Key innovations include perception-driven phase segmentation, arm–hand协同 compensation for pose errors, and a balanced adaptive force coordination mechanism. By analytically linking contact forces to joint motions, the method remains compatible with diverse grasp pose generation strategies. Evaluated across 15,000 simulations involving 478 objects and eight physical experiments, the approach significantly improves grasp success rates while effectively suppressing unintended object motion.

This work addresses the challenge of directly transferring human grasping motions to dexterous robotic hands, which often fails due to morphological and contact constraints. To overcome this, the authors propose a novel approach that integrates functional-aware human pre-grasp synthesis with robot-native contact optimization. The method leverages object-conditioned digital human pre-grasp sampling, hand pose retargeting, and force-closure-based contact refinement, augmented by a vision-language model (VLM) agent for task-level planning to generate high-quality, anthropomorphic end-to-end manipulation demonstrations. Evaluated in simulation, the approach achieves an 80.7% success rate; on a real-world 36-degree-of-freedom bimanual robot platform, it completes 25 out of 30 tasks (83.3%) with 86.4% grasp stability and 93.4% anthropomorphic fidelity, marking the first effective fusion of human-inspired pre-grasps and robot-native contact optimization.