Grasp planning is the computational problem of determining stable contact configurations between a robotic end-effector and an object, such that the resulting grasp resists external disturbances and enables the desired manipulation task. It combines geometric modelling of object shape, force-closure analysis, kinematics constraints, and task-level objectives to synthesise executable grasp poses.
Content
- Grasp planning research emerged in the 1980s alongside the development of robotic manipulators for manufacturing. Pioneering work by Mason and Salisbury (1985) formalised force-closure conditions — the requirement that contact forces span the space of external wrenches. The 1990s brought analytical approaches to grasp quality metrics (Ferrari and Canny’s 1992 ε-metric), and the 2000s saw Monte Carlo and sampling-based planners (GraspIt!, OpenRAVE) that could handle complex 3D geometry.
- Grasp planning algorithms fall into three broad categories: analytical methods (compute force-closure configurations from object geometry and contact models), sampling-based methods (sample candidate grasps, evaluate quality, select the best), and data-driven methods (learn grasp success from large datasets or simulation rollouts). A typical pipeline proceeds as: acquire 3D point cloud of the object, estimate object pose (6-DoF), generate grasp candidates (via geometrical primitives or neural sampling), evaluate each candidate for stability and task compatibility, select the highest-quality collision-free grasp, compute approach trajectory via motion planning, and execute with force feedback.
- Grasp planning is essential for industrial pick-and-place operations (logistics, assembly), surgical robotics (instrument handling), household service robots (domestic manipulation), and agricultural harvesting. The accuracy and robustness of grasp planning directly determines a robot’s ability to operate in unstructured environments where object placement, orientation, and surface properties are unpredictable. Real-world deployment is complicated by sensor noise, deformable objects, clutter, occlusion, and the sim-to-real gap.
- In 2024-2025, large-scale generalist grasp models trained on millions of simulated and real grasps (GraspNet-1Billion, AnyGrasp, DexGraspNet) have substantially advanced zero-shot generalisation to novel objects. Foundation models applied to robotics (RT-2, OpenVLA) encode grasp planning implicitly within end-to-end visuomotor policies. Tactile sensing integration for in-hand adjustment and slippage detection is maturing, and sim-to-real transfer via domain randomisation and physics-based rendering now enables training entirely in simulation with strong real-world transfer.