Event
GRASP: A Global-Local Representation Alignment Framework for Dual-Arm Manipulation
1 reports · 1 sources
Zeyuan Ren et al. propose GRASP, a training-time prior injection framework for fine dual-arm robotic manipulation. It supervises tool center point poses and inter-arm relative poses via a physically anchored flow matching loss through differentiable forward kinematics, and introduces a global query group distilled from DINOv3 and local interaction queries supervised by Grounding DINO; both teacher models are removed at inference time, and the deployment pipeline remains a standard joint-space flow policy with only the learned query pathways added. Simulation and real-world long-horizon dual-arm task experiments show that the two priors are complementary under different deployment conditions: the full model performs better in in-distribution deployment, while the distilled query branch is more applicable under strong visual randomization.
Report timeline
- GRASP: Global-Local Representation Alignment and Spatial-Physical Grounding for Bimanual Robotic Manipulation
GRASP proposes a training-phase prior injection framework for bimanual robotic manipulation. A physics-grounded flow-matching loss uses differentiable forward kinematics (FK) to supervise tool center point (TCP) poses and inter-arm relative poses, and the framework introduces a global query group distilled from DINOv3 together with local interaction queries supervised by Grounding DINO; the teacher models are removed at inference. Simulation and real-world long-horizon bimanual task experiments show that the two priors are complementary under different deployment conditions: the full model performs better in in-distribution deployment, while the distilled query branch is more suitable under strong visual randomization.
arXiv cs.RO ·