What is 3D point cloud data collection?
3D point cloud data collection is the capture of spatial geometry as sets of XYZ points, often with color, normals, and intensity, using depth cameras, stereo rigs, structured light, photogrammetry, or laser scanners. For robotics, the goal is geometry a model can trust: complete enough to grasp against, accurate enough to register, and stored with the sensor and pose data behind it.
This page covers near-field and object-level geometry: the clouds behind grasp prediction, bin picking, digital twins of work cells, and simulation assets. If you're capturing long-range scans from a moving robot or vehicle, LiDAR data collection is the better fit.
Where point cloud data breaks: glass, shine, and thin parts
Commodity depth sensors are good at matte, textured, medium-sized objects. They struggle with exactly the objects that show up in kitchens, labs, and warehouses: glass, polished metal, black plastic, and anything thin.
Google's ClearGrasp work measured what that costs a robot. When raw depth on transparent objects was replaced with corrected depth, grasp success rose from 12% to 74% with a parallel-jaw gripper and from 64% to 86% with suction.
The robot was never the problem. The geometry it was given was.
Picture a home-robot team testing a dish-loading policy. It performs well on ceramic plates, then fails on every wine glass and stainless pot. Nobody changed the model. The training clouds simply never contained honest geometry for those materials, only holes and noise that got smoothed into shapes that don't exist.
Choosing a 3D capture method
We pick the method from the object set and the downstream use. Most programs combine two: a fast method that matches what the robot sees live, and a precise one that serves as reference geometry.
| Method | Best for | Strengths | Fails on |
|---|---|---|---|
| Active stereo RGB-D | Grasping and bin picking, matching live robot perception | Fast, color-aligned, inexpensive | Transparent, black, and specular surfaces; noisy edges |
| Structured light | Precise object scans, inspection | High accuracy at close range | Sunlight and moving scenes |
| Time-of-flight cameras | Mid-range, dynamic scenes | Works on low-texture surfaces | Multipath near corners; flying pixels at edges |
| Photogrammetry | Textured objects, large scenes, photoreal sim assets | Rich color, scales cheaply | Textureless and shiny surfaces; needs a scale reference |
| Turntable multi-view scanning | Object libraries and asset creation | Complete coverage including undersides | Articulated or deformable objects without extra passes |
| Terrestrial and handheld laser scanners | Rooms, work cells, facilities | Range and accuracy | Fine detail on small objects |
The Cloud Fitness Score: five checks before a cloud ships
Gamasome framework
A point cloud can look complete on screen and still be unfit for training. We score every cloud on five dimensions before it enters your dataset, and we report the scores alongside the data so your team can filter by them.
Filled holes are the most dangerous data in a point cloud. They look right and they're made up.
Coverage
The share of the object's surface actually observed, measured against a reference mesh or the union of views. Undersides, interiors, and handles are where most scans fall short.
Noise floor
Residuals from fitting known planes and primitives in the scene. It tells you whether edges and thin parts are real geometry or sensor noise.
Registration error
How tightly multiple views align. Drift here doubles surfaces and quietly shifts grasp poses by millimeters that matter.
Material honesty
Transparent, specular, and dark regions are flagged, not silently filled. Your model should know where the sensor was guessing.
Provenance
Sensor model, intrinsics, extrinsics, pose, and timestamp stored with every cloud, so any cloud can be traced back to how it was captured.
From scan to simulation-ready asset
A growing share of point cloud work ends in simulation. That's where a visually perfect asset can be physically useless: wrong mass, missing joints, collision meshes that explode or interpenetrate the moment the robot touches them.
Turing: sim-ready household assets for NVIDIA Isaac Sim
- Situation
- Turing needed household 3D assets converted into objects that behave correctly in NVIDIA Isaac Sim.
- Problem
- Assets that looked right broke in simulation because physics properties, joints, and collision geometry were missing or wrong.
- Solution
- Gamasome built a USD authoring pipeline (on OpenUSD) that adds correct physics, joints, and collision behavior, with a validation standard every asset must pass.
- Outcome
- 70 of 100 target assets delivered as of July 2026, each validated against the same repeatable standard.
For a robotics team, the practical win is that a drawer in simulation opens like a drawer, and a mug falls like a mug. Sim-to-real testing only means something when the sim objects obey the same physics as the real ones.
Where teams use our point cloud data
- Grasp prediction and bin-picking datasets with cluttered, realistic scenes
- Digital twins of work cells for layout, reach, and collision checks
- Articulated object capture (drawers, doors, lids) with joint states recorded
- Reference geometry for evaluating depth completion and 3D perception models
- Source geometry for sim-ready assets and synthetic data generation
What teams get wrong with 3D capture
Meshing too early
Once a cloud becomes a mesh, the evidence of where the sensor guessed is gone. Keep raw clouds alongside every derived mesh.
Photogrammetry with no scale reference
Photogrammetry has no inherent scale. Without scale bars or known markers in frame, a mug can come out the size of a bucket.
Hero poses on clean backgrounds
Robots see objects in clutter, at odd angles, partly hidden. Object libraries need scene captures, not just studio scans.
Dropping articulation
A cabinet scanned closed can't teach a robot to open it. Articulated objects need multiple states and recorded joint positions.
Formats and handoff
Clouds ship as PLY, PCD, E57, or LAS/LAZ, meshes as OBJ or GLB, and sim assets as USD with physics, alongside intrinsics, extrinsics, poses, and Cloud Fitness Scores. Our data annotation team handles 3D segmentation and grasp labels. For depth paired with force and touch during manipulation, see multimodal data collection.
3D point cloud data collection FAQs
What is 3D point cloud data collection?
It's the capture of 3D geometry as sets of points, often with color and normals, using depth cameras, stereo, structured light, photogrammetry, or scanners. For robotics, the data trains and evaluates grasping, 3D perception, and scene understanding, and feeds simulation assets.
What's the difference between point cloud and LiDAR data collection?
LiDAR data collection focuses on long-range scans from moving robots and vehicles for perception, mapping, and localization. Point cloud collection covers near-field and object-level geometry for grasping, digital twins, and simulation assets, using a wider range of 3D sensors.
How do you capture transparent or reflective objects?
We combine capture methods, use reference geometry from precise scans, and flag unreliable regions rather than filling them. Research such as Google's ClearGrasp shows corrected depth on transparent objects can raise grasp success from 12% to 74%, so honest geometry here matters.
Can you turn scans into simulation-ready assets for Isaac Sim?
Yes. We build USD assets with physics properties, joints, and collision geometry validated against a repeatable standard, as in our Turing program for NVIDIA Isaac Sim.
Which point cloud formats do you deliver?
PLY, PCD, E57, and LAS or LAZ for clouds, OBJ or GLB for meshes, and USD for simulation assets, with intrinsics, extrinsics, poses, and quality scores included.
Do you also label point clouds?
Yes. Our data annotation team handles 3D segmentation, part labels, bounding boxes, and grasp annotations, working from the same specification as the capture plan.