6-DoF Grasp Pose

• Also called: 6-DoF grasp
• Category: Planning & manipulation › Manipulation
• Tags: grasping, pose-estimation, point-clouds, manipulation, grasp-detection

The full 3D position and 3D orientation of a gripper at which it should close on an object, as opposed to planar or top-down grasps with fewer free parameters.

Definition

A 6-DoF grasp pose specifies where a gripper should be when it closes on an object: three degrees of freedom for position and three for orientation, together an element of SE(3). It is usually expressed as the pose of a gripper reference frame, often at the center between the fingertips, relative to the camera, the robot base or the object. For a parallel-jaw gripper the description is often extended with the opening width, and for multi-fingered hands with finger joint angles, but the 6-DoF part is what lets the gripper approach from any direction.

The term is used mainly in contrast with lower-dimensional grasp representations. Early learning-based grasping predicted planar grasps, such as oriented rectangles in an image, which amount to a top-down approach with a position, a depth and a rotation about the vertical axis, about four degrees of freedom. These are adequate for objects lying flat in a bin viewed from above, but they cannot grasp a mug from the side, a bottle lying on a shelf or an object held by a person. 6-DoF methods search the full space of approach directions, which makes them more general but also harder: the space is larger, many candidate poses collide with the object or the scene, and the result must be reachable by the arm.

Typical 6-DoF pipelines start from a partial point cloud. Sampling-based methods such as Grasp Pose Detection propose many gripper poses using geometric heuristics like antipodal contact and then score them with a learned classifier. Generative methods such as 6-DOF GraspNet sample grasp poses from a learned distribution and refine them with a learned evaluator, while Contact-GraspNet predicts grasps anchored at observed surface points of a full scene. Each candidate is then checked for collisions, solved with inverse kinematics, and executed with a pre-grasp pose a short distance back along the approach axis, a straight-line approach, closing and lifting. Frame conventions matter: libraries differ in which gripper axis is the approach direction and where the frame origin sits, and a parallel-jaw grasp rotated 180 degrees about its approach axis is physically the same grasp, which affects both training losses and evaluation.

Why it matters

Most real objects are not lying flat in a bin. Shelves, tables, clutter and objects held by people all require approaching from directions other than straight down, and 6-DoF grasp poses are the representation that makes those grasps possible. They are also the common interface between grasp detectors and motion planners.

Key applications

• Picking objects from shelves, tables and cluttered scenes
• Human-to-robot handovers of objects held in a person's hand
• Mobile manipulation in homes and warehouses
• Grasping previously unseen objects from partial point clouds
• Generating grasp datasets in simulation for learning

In practice

A kitchen robot needs to pick a mug standing upright with its handle facing away. A depth camera gives a partial point cloud of the mug, and a grasp network returns a few hundred 6-DoF grasp candidates with scores. The software transforms the candidates into the robot base frame, drops those that collide with the table or the scene, and runs inverse kinematics on the rest, keeping those that are reachable without approaching a joint limit. For the best grasp, which closes on the rim from the side, the planner moves to a pre-grasp pose 10 centimeters back along the gripper's approach axis, approaches in a straight line, closes the gripper and lifts. An early bug traced to the grasp library and the robot's gripper using different approach axes, which was fixed with a fixed rotation between the two frame conventions.

Common misconceptions

• A 6-DoF grasp pose is enough to execute a grasp — it must still pass collision checks, have an inverse-kinematics solution and be reached through a sensible pre-grasp approach
• Top-down grasps are enough for most tasks — they work for objects lying flat in open bins, but shelves, side-facing handles and objects held by people need approach directions a top-down grasp cannot provide
• All grasp libraries use the same gripper frame — the approach axis and frame origin differ between tools, and mixing conventions silently produces wrong grasps
• 'GraspNet' names one method — 6-DOF GraspNet and the GraspNet-1Billion dataset and baseline are separate works from different groups

Who should know this

• Manipulation and grasping engineers
• Perception engineers building pick-and-place systems
• Robot learning researchers working on grasp detection
• Mobile manipulation developers

Related technologies

• 6-DOF GraspNet
• Contact-GraspNet
• Grasp Pose Detection (GPD)
• MoveIt
• Open3D

Related terms

• Grasp Planning
• Antipodal Grasp
• Grasp Quality Metric
• SE(3)
• Point Cloud
• 6D rotation representation

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