6-DOF Grasp Planning: Definition, Applications, and Training Data Requirements

6-DOF grasp planning is the computational problem of determining a full six-dimensional gripper pose — three translational coordinates (x, y, z) and three rotational angles (roll, pitch, yaw) — that will result in a stable grasp on a target object. The "6-DOF" refers to the six degrees of freedom in the rigid-body transformation group SE(3), which fully specifies the position and orientation of the end-effector in three-dimensional space. Unlike planar or top-down grasping approaches that restrict the gripper to approach from above, 6-DOF planning enables robots to grasp objects from any direction, which is essential for real-world tasks involving shelving, bin-picking, and cluttered tabletop manipulation.

The canonical 6-DOF grasp planning pipeline begins with perception: a depth camera or LiDAR sensor produces a point cloud or depth image of the workspace. A grasp proposal network then generates candidate grasp poses — typically hundreds or thousands per scene — scored by a predicted probability of grasp success. The highest-scoring collision-free grasp is selected and sent to a motion planner that computes a feasible arm trajectory to the pre-grasp position, followed by approach and closure. Analytical methods based on force-closure analysis dominated early work, but learning-based approaches now achieve significantly higher success rates on novel objects by training on large datasets of grasp attempts.

Modern grasp planners fall into two broad families. Sampling-based methods like GPD (Grasp Pose Detection) generate grasp candidates by sampling points on the object surface and evaluating each with a learned classifier. Regression-based methods like Contact-GraspNet directly predict per-point grasp parameters from the input point cloud, avoiding the expensive sampling loop. A third emerging category uses diffusion models or score-based generative models to sample grasps from a learned distribution over SE(3), enabling multi-modal grasp prediction where several distinct grasp strategies (power grasp, pinch grasp, side grasp) are proposed for the same object.

The quality of training data is the primary determinant of 6-DOF planner performance. Grasp annotations must include the full SE(3) pose, gripper width, and a binary or continuous success label. Datasets must cover diverse object geometries — from convex boxes to concave bowls to articulated tools — and scene configurations ranging from isolated objects to dense clutter with significant occlusion. Without this diversity, planners overfit to training object shapes and fail on novel items encountered during deployment.

For teams building production grasping systems, the first design decision is the perception modality. Most 6-DOF planners operate on point clouds from commodity depth sensors — the Intel RealSense D415/D435 and Microsoft Azure Kinect are the most common in research. Sensor placement matters significantly: eye-in-hand (camera mounted on the wrist) provides close-up views during approach but limited global scene awareness, while fixed overhead or shoulder-mounted cameras give broader context but lower spatial resolution on small objects. Many production systems use both configurations, fusing point clouds from two or three viewpoints.

Data collection for 6-DOF grasping follows one of three paradigms. First, pure simulation: generating millions of grasp labels in Isaac Sim or MuJoCo against procedurally generated or scanned object meshes, then transferring to reality with domain randomization. Second, self-supervised physical collection: the robot autonomously attempts grasps and records binary success outcomes, as in the Google brain grasping project that collected 800,000 grasp attempts across 7 robots. Third, human annotation: experts label grasp poses on point cloud visualizations, which produces higher-quality labels but scales poorly. The practical optimum combines simulated pretraining with 5,000-20,000 real-world grasp executions for domain adaptation.

Evaluation must happen in the real world. Offline metrics on held-out scenes do not reliably predict physical success rates because they cannot capture sensor noise, calibration error, or the dynamics of contact. Standard physical benchmarks use 20-50 novel objects arranged in bins at varying clutter densities. Success rate (percentage of grasps that lift and hold the object for 3+ seconds) and clearance rate (percentage of objects cleared from a bin in N attempts) are the metrics that matter for deployment.

Claru supports 6-DOF grasping teams by providing annotated point cloud datasets from real warehouse, kitchen, and industrial scenes with ground-truth depth maps, instance segmentation masks, and object pose labels. These datasets serve both as training data for grasp planners and as standardized evaluation sets for comparing system performance across object categories and clutter densities.

Claru provides the real-world perception data that 6-DOF grasp planners need for training and evaluation. Our catalog includes RGB-D point cloud captures from warehouse, kitchen, and industrial environments with dense clutter, partial occlusions, and diverse object geometries — the exact conditions where production grasping systems must operate. Each scene includes calibrated depth maps, instance segmentation masks, and object pose annotations that enable teams to generate ground-truth grasp labels for their specific gripper geometry.

For teams transitioning from simulated to real-world grasping, Claru offers targeted data collection: high-resolution depth scans of customer-specified object categories captured under varied lighting, sensor angles, and background conditions. Our 10,000+ data collectors across 100+ cities can capture the environmental diversity that prevents grasp planners from overfitting to a single lab setup. The data passes through a quality pipeline that validates depth completeness, point cloud registration accuracy, and annotation consistency — addressing the data quality issues that cause the largest performance gaps between simulated and real-world grasping.