Grasp planning is the process of deciding where and how a robot gripper or hand should contact an object so that it can be held securely and moved for a task. In robotic manipulation, it produces a grasp pose, meaning the position and orientation of the gripper, together with the finger configuration and, in some cases, the required grip force.
A grasp planner must balance three demands: the grasp must be physically stable, it must be reachable without collisions, and it must suit the next step of the task, such as placing, pouring or inserting the object. The planner therefore sits between perception, which describes the object, and motion planning, which moves the arm to the chosen pose.
How grasp planning works
Most grasp planners follow a generate, evaluate and select pattern. They first produce many candidate grasps around the object, then score each candidate for quality and feasibility, and finally pass the best one to the motion planner. The quality of the result depends as much on the candidate generator as on the scoring.
- Perceive the object as a known model with an estimated pose, or as a partial point cloud of an unknown shape.
- Sample candidate grasps, for example antipodal finger placements or approach directions along surface normals.
- Score each candidate with an analytic quality metric or a learned success predictor.
- Check feasibility with inverse kinematics and collision checking for the arm and gripper.
- Select and execute the highest-ranked feasible grasp, with pre-grasp and retreat poses.
The feasibility check links grasp planning to inverse kinematics: a grasp that is perfect for the object is useless if the arm cannot reach it within joint limits or without colliding with the bin walls.
Grasp quality and stability
Analytic grasp planning relies on contact mechanics. Each finger contact can transmit forces inside a friction cone whose half-angle depends on the friction coefficient. A grasp achieves force closure when the contacts together can resist any external force and torque, and form closure when geometry alone prevents motion, even without friction.
Planners compare grasps with numeric metrics. The widely used Ferrari-Canny epsilon metric measures the largest disturbance wrench that the grasp can resist in the worst direction, given unit contact forces.
| Criterion | What it measures | Typical use | Limitation |
|---|---|---|---|
| Force closure | Ability to resist any wrench | Binary feasibility test | Says nothing about margin |
| Epsilon metric | Worst-case wrench resistance | Ranking analytic grasps | Sensitive to friction estimate |
| Antipodal score | Alignment of opposing contact normals | Parallel-jaw grippers | Ignores object mass distribution |
| Distance to center of mass | Torque induced by gravity | Heavy or long objects | Center of mass often unknown |
| Learned success probability | Predicted chance of a stable lift | Unknown objects from images | Depends on training data coverage |
Friction coefficients for rubber fingertips on common materials often lie roughly between 0.3 and 1.0, and planners usually assume a conservative value because an overestimate yields grasps that slip in practice.
Analytic, data-driven and learned approaches
Approaches differ mainly in how much they know about the object beforehand. When an exact CAD model is available, grasps can be precomputed offline and simply transformed by the estimated object pose. For unknown objects, the planner must reason directly from a partial view.
| Approach | Input | Strength | Weakness |
|---|---|---|---|
| Analytic, model-based | Known 3D model and pose | Physically grounded, explainable | Needs accurate models and pose |
| Grasp database | Object identity and pose | Very fast at runtime | Limited to cataloged objects |
| Geometric heuristics | Point cloud or depth image | Works on novel shapes | Ignores friction and dynamics |
| Learned grasp detection | RGB or depth images | Generalizes to clutter | Needs large training datasets |
| Reinforcement learning | Sensor stream and reward | Can learn regrasping and pushing | Costly training, sim-to-real gap |
In industry, hybrid systems are common: a learned model proposes candidates in clutter, and analytic checks reject grasps that violate reachability or collision constraints.
Choosing a planner for the gripper and task
The end effector largely determines the planning problem. A parallel-jaw gripper needs only a pose and an opening width, a suction cup needs a flat and sealed surface patch, and a multi-fingered hand adds many joint angles and contact choices.
| End effector | Decision variables | Key planning concern |
|---|---|---|
| Parallel-jaw gripper | 6D pose plus width | Antipodal contacts within stroke |
| Vacuum suction cup | Contact point and approach axis | Flat, nonporous surface near center of mass |
| Multi-fingered hand | Pose plus finger joint angles | High-dimensional search, contact stability |
| Soft or adaptive gripper | Approach pose | Compliance tolerates pose error |
The task matters as much as the gripper. In pick and place, a grasp that is stable for lifting may block the final placement, so a task-aware planner considers the place pose and may plan a regrasp.
Common failure modes
Grasp failures rarely come from the scoring formula alone. They usually arise where perception, mechanics and motion meet.
- Pose error: a few millimeters of pose error can turn a centered grasp into an edge grasp that slips.
- Occluded geometry: the planner sees only the visible side and misjudges the object’s width or depth.
- Collision with clutter: the fingers fit around the object but hit neighboring items on approach.
- Unmodeled mass: an off-center grasp on a heavy object rotates it within the fingers.
- Transparent or shiny objects: depth sensors return missing data, so candidates are placed on phantom surfaces.
Understanding it simply
When a person picks up a mug, they decide without thinking whether to grab the handle or wrap a hand around the body. A robot has to make that decision deliberately. Grasp planning is how it chooses where to put its fingers and from which direction to approach.
A good choice holds the object firmly, can be reached without bumping into anything, and leaves the object ready for the next step. A grasp is only good if the whole task succeeds, not just the lift.
Applied in practice
An e-commerce fulfillment cell must pick thousands of different products, from boxed electronics to soft bags, out of mixed totes. The system uses a depth camera above the tote and a learned network that proposes hundreds of suction and parallel-jaw grasp candidates per image.
Each candidate is filtered by an inverse kinematics check and a collision check against the tote walls, and the remaining ones are ranked by predicted success. Suction is preferred for flat boxes, while the gripper handles irregular items. When a pick fails, the image and outcome are stored, so the network is retrained on real failure cases and the pick rate steadily improves.