We address the problem of selecting feasible 6-DoF grasp poses from a pre-trained SE(3)-equivariant generative model without any additional training. While recent grasp generators based on diffusion or flow matching can sample diverse grasp candidates...
We address the problem of selecting feasible 6-DoF grasp poses from a pre-trained SE(3)-equivariant generative model without any additional training. While recent grasp generators based on diffusion or flow matching can sample diverse grasp candidates for a given object, they typically rely on fixed, learned grasp-quality criteria and two-stage pipelines that first generate many candidates and then post-select a few feasible grasps. We propose GGPS (Guided Grasp Pose Sampler), a training-free test-time guidance method that directly samples scenario-aware feasible grasps on SE(3). Given a user-defined energy function \(J : \mathrm{SE}(3) \to \mathbb{R}\) that evaluates grasp poses under the current robot, environment, and task, GGPS modifies the flow dynamics only at sampling time so that the induced distribution concentrates on low-energy regions. Building on recent work on guidance for flow matching, we extend their Euclidean formulation to matrix Lie groups. Instantiated on EquiGraspFlow, our approach enables the same pre-trained model to realize different notions of feasibility—incorporating affordance, reachability, manipulability, or task-specific efficiency—simply by changing \(J\), without any retraining or additional optimization.