grasp synthesis

This work addresses the semantic–physical gap between high-level visual-language understanding and low-level robotic manipulation, enabling zero-shot generalization to novel tasks in unstructured environments without task-specific training. Method: We propose Executable Abstraction Concepts (EACs), a unified mathematical formalism that jointly encodes object functionality, geometric constraints, and operational semantics—thereby establishing an interpretable, end-to-end mapping from language instructions to executable robot actions. Our framework integrates vision-language models (VLMs), natural language processing (NLP), geometric reasoning, and motion planning to generate grasp poses, force directions, and dynamically feasible trajectories. Contribution/Results: Evaluated in both simulation and real-world settings, our approach achieves strong zero-shot generalization across diverse articulated objects under complex natural-language commands, significantly improving manipulation accuracy and environmental adaptability compared to prior methods.

This work addresses the fragmentation between symbolic planning and geometric motion planning in language-to-robotic-manipulation trajectory generation. We propose a robot-agnostic, planner-agnostic, and task-agnostic language-driven Task and Motion Planning (TAMP) paradigm. Methodologically, we introduce the first plug-and-play integration of large language models (LLMs) with the Kautham motion planning framework: natural language instructions are parsed into executable action sequences via symbolic logic reasoning, then coupled with collision-aware, kinematically and dynamically constrained trajectory optimization for end-to-end trajectory synthesis. Evaluated in a ROS-compatible robotic arm simulation environment, our system enables zero-code adaptation across multi-arm setups and diverse scenarios. It significantly improves the success rate of instruction-to-executable-trajectory translation and cross-task generalization. To our knowledge, this is the first fully stackable, physically grounded robotic manipulation planning framework driven entirely by LLMs.

Constructing high-quality, cross-hardware datasets for contact-rich robotic manipulation remains challenging due to high data acquisition costs and hardware-specific constraints. Method: This paper proposes a low-cost, high-fidelity multi-source data generation framework. It introduces the first “embodiment-flexible” VR-based human demonstration capture system, integrated with physics simulation, kinematic retargeting, and constraint-aware, multi-body parameter-adaptive trajectory optimization—enabling cross-configuration data reuse and zero-shot transfer. Contribution/Results: Using this pipeline, we build a large-scale contact manipulation dataset and train a diffusion-based policy with strong generalization. Evaluated on Allegro Hand and dual-arm iiwa platforms, the policy achieves zero-shot deployment on real dual-arm iiwa hardware, significantly improving success rates on contact-intensive tasks with minimal human intervention. Key contributions are the embodiment-flexible data collection paradigm and the physically consistent, parameter-adaptive optimization mechanism.

This work addresses the prevalence of physically implausible artifacts—such as interpenetration, missing contacts, and unstable grasps—in existing hand-object interaction datasets generated via physics simulation. To resolve these issues, the authors propose a closed-loop simulation-based optimization framework that transforms visually aligned hand-object trajectories into physically executable motions. The approach parameterizes actions using low-dimensional splines anchored at sparse keyframes and employs a high-fidelity physics engine as a black-box objective function. Gradient-free optimization via CMA-ES is used to refine the trajectories, ensuring physical plausibility while preserving fidelity to the original motion. Compared to prior methods such as MANIPTRANS, this technique significantly reduces hand-object pose replay error and more accurately recovers complex interactive behaviors, including stable grasping and lifting.

To address the trade-off between computational efficiency and modeling expressiveness in dynamic modeling of high-dimensional deformable objects, this paper proposes a task-driven, spatially adaptive method for automatic dynamic model generation. The method introduces diffusion models to predict local modeling resolution at critical regions from point-cloud-based planning queries—a novel application in this domain. It further designs a two-stage joint optimization framework that unifies predicted dynamics priors with closed-loop control performance feedback, enabling co-optimization of spatial resolution distribution and dynamical parameters. Integrating point-cloud perception, differentiable physics simulation, and closed-loop simulation data collection, the approach achieves a 2× speedup over full-resolution modeling on tree-like object manipulation tasks, with less than 3% degradation in task success rate. This substantially enhances both planning efficiency and practical deployability.

This work addresses the tight coupling between geometric planning and motion execution in existing robotic surface interaction methods, as well as the limited generalization of imitation-learned policies to novel geometries. To overcome these challenges, the authors propose a modular framework that decouples geometric path planning from expert-level motion execution. The approach dynamically adapts reference trajectories through interpretable atomic motion rules—such as velocity scaling and pose offsetting—and employs a multimodal neural network to jointly learn task trajectories and CAD-based geometric features, thereby inferring rule parameters. This enables, for the first time, a disentangled representation of geometry-aware perception and transferable motor skills. Experiments in dynamic simulation on L-shaped and window-shaped objects successfully extract generalizable motion rules, demonstrating the method’s strong generalization across topologically distinct tasks.

Existing approaches struggle to autonomously generate diverse sequences of contact locations and manipulation trajectories, limiting their ability to perform complex, contact-intensive tasks. This work proposes SCSP, a cascaded optimization framework that unifies contact location selection and trajectory planning within a single online-executable optimization pipeline. By integrating a surrogate contact model, mixed discrete-continuous optimization, and prior-guided real-time trajectory generation, SCSP effectively addresses the challenges posed by complementarity in contact dynamics and sparse gradients. Implemented on redundant robotic arms, the method enables joint online synthesis of contact points and motion trajectories. Extensive simulations and real-world experiments demonstrate its capability to produce diverse manipulation behaviors, exhibit robustness against dynamic modeling errors and perception noise, and generalize effectively across complex contact-rich tasks.

This work addresses the challenges in non-prehensile dexterous manipulation—namely, infeasible actions due to neglecting gripper physical constraints, poor generalization, and heavy reliance on large datasets or manual design—by proposing a gripper-aware hierarchical planning framework. Centered on executability, the approach decouples object motion planning from grasp feasibility: an upper layer plans object pose trajectories using MoveObject primitives, while a lower layer synthesizes feasible grasp sequences via AdjustGrasp primitives, with collision checking and quasi-static force analysis validating contact-sensitive segments. The method requires no task-specific redesign and supports transfer across tasks and geometric variations. Real-robot experiments on zero-displacement lifting and slot-insertion tasks demonstrate strong robustness and consistent execution performance.

This work addresses the challenge of effective exploration in contact-rich, sparsely rewarded environments where general-purpose reinforcement learning struggles to discover complex manipulation strategies. The authors propose a Sample-Guided RL framework that leverages a differentiable model-based solver—incorporating collision, contact, and force constraints—to construct a low-dimensional manifold of feasible states and guide policy learning through targeted sampling. By integrating black-box optimization to generate open-loop trajectories and introducing state-visit bias alongside behavioral cloning loss, the method significantly enhances both training efficiency and performance of goal-conditioned policies. Evaluated on a simplified two-sphere environment and Panda robot arm tasks, the approach substantially outperforms baseline methods, achieving high success rates in reaching statically stable states and demonstrating diverse whole-body contact-aware manipulation strategies.

This work proposes a novel framework for inverse kinematics (IK) optimization that addresses the high failure rates commonly caused by the nonlinear relationship between joint variables and end-effector poses, as well as non-convex constraints such as obstacle avoidance. By introducing analytical IK solutions as a change of variables within the optimization process, the method uniquely combines the precision of analytical approaches with the flexibility of numerical optimization, substantially simplifying the problem structure. Evaluated across three mainstream optimizers, the approach demonstrates significantly higher success rates than conventional optimization techniques and baseline methods in complex tasks—including obstacle avoidance, grasp selection, and humanoid robot stability—thereby achieving an effective unification of analytical and optimization-based IK strategies.