Antipodal Grasp
- Also called: antipodal point grasp
- Category: Planning & manipulation › Manipulation
- Tags: grasping, parallel-jaw, point-clouds, manipulation, contact
A two-contact grasp in which the line joining the contacts lies inside the friction cone at each contact, so the fingers can squeeze the object without slipping.
Definition
An antipodal grasp places two contacts on opposite sides of an object so that the line between them is aligned with the surface normals at both points. In the strict geometric sense the two normals point exactly at each other along that line. With friction the condition relaxes: the grasp is antipodal when the line joining the contacts lies within the friction cone at each contact, which means the squeezing force can be transmitted without either finger slipping. This is the shape of grasp a parallel-jaw gripper makes, which is why the term is so common in practical grasping work.
For planar objects, a frictional antipodal grasp is force closure: it can resist any in-plane disturbance given enough squeeze. In 3D the picture is less complete. Two point contacts cannot resist torque about the axis between them, so a real antipodal grasp also relies on torsional friction from the finite contact patches of rubber or soft finger pads. Gravity acting on an object held far from its center of mass creates exactly such a torque, so long objects gripped at one end can rotate out of an otherwise sound antipodal grasp.
Antipodality is cheap to check, which makes it a workhorse in grasp sampling and labeling. Given a point cloud with estimated normals, a sampler can propose a gripper pose, find the points that would be touched by each jaw and test whether their normals lie within a friction-cone angle of the closing direction. Grasp Pose Detection and Dex-Net use this kind of geometric check to generate or score large numbers of candidate grasps, and learned grasp networks are often trained on labels produced this way.
Why it matters
Most deployed robot grippers are parallel-jaw, and an antipodal check is the simplest geometric test of whether such a gripper will hold an object. It turns grasp planning on raw point clouds into a fast sampling problem and supplies labels for learned grasp detectors.
Key applications
- Grasp sampling for parallel-jaw grippers on point clouds
- Generating training labels for learned grasp detectors
- Bin picking and warehouse order picking
- Grasp candidate filtering before motion planning
- Benchmarking grasp quality in simulation
In practice
A bin-picking cell uses a wrist camera to capture a point cloud of mixed parts. The grasp module estimates surface normals, then samples gripper poses whose closing axis is roughly aligned with local normals. For each pose it checks that the gripper body does not collide with the cloud, finds the points between the jaws, and accepts the grasp if points on both sides have normals within about 15 to 20 degrees of the closing direction, a margin chosen below the friction-cone angle of the rubber finger pads. Accepted grasps are ranked by a learned score and by how close the jaw center is to the part's estimated centroid, which reduces rotation caused by gravity torque about the jaw axis.
Common misconceptions
- Antipodal means the normals must be exactly opposite — with friction, the line between contacts only needs to lie inside both friction cones
- An antipodal grasp is force closure in 3D — two point contacts leave rotation about the jaw axis unresisted, so torsional friction from the pads is also required
- Passing the antipodal check guarantees a successful pick — mass distribution, collisions, normal estimation noise and object deformation can still cause failure
Who should know this
- Engineers building pick-and-place and bin-picking systems
- Grasp detection and learning researchers
- Gripper and finger-pad designers
- Robotics students studying grasp mechanics
Related technologies
- Parallel-Jaw Gripper
- Grasp Pose Detection (GPD)
- Dex-Net
- Surface Normal Estimation
- Point Cloud Library
Related terms
Related projects
- LeRobot: HuggingFace ML library for real-world robotics in PyTorch
- Isaac Lab: NVIDIA robot learning framework: one task API across Newton, PhysX and Isaac Sim
- NVIDIA Isaac GR00T: NVIDIA open foundation model for humanoid robot reasoning and skills
- OpenVLA: 7B-parameter Vision-Language-Action model for robotic manipulation
- Drake: MIT robotics toolbox for model-based dynamics, optimization, and control system design
- OpenArm: 7DOF humanoid arm for physical AI research in contact-rich environments