3D & Spatial Annotation for Robot Manipulation
Grasp poses, affordance regions, 6-DoF object pose, part structure and articulation—the annotations a robot needs to interact rather than merely detect.
Perception Tells You What Is There. Manipulation Needs How.
For perception cuboids and scene tracks, see 3D point cloud & LiDAR annotation.
What We Annotate
- 6-DoF object pose
- Task-conditioned grasp pose sets
- Affordance regions
- Functional part segmentation
- Articulation and physical properties
There Is No Single Correct Grasp
We measure coverage, correct exclusion, task conditioning, gripper specificity and robustness.
Verifying Automated Annotation
Published automatic-annotation research reported 26.6% physics validation and 23.2% final retention on its own audited suite. We verify candidates and report generator failure patterns.
Physics Validation Is Not Real-World Validation
Simulation is a filter. Hardware evidence is separately measured when explicitly in scope.
How a Programme Runs
- Scope and guideline
- Asset preparation
- Annotator calibration
- Pilot batch
- Production or verification
- Physical verification
- Guideline iteration
Related Services
Human Demonstrations Multimodal Sensor Data Image Annotation
3D Spatial Annotation FAQs
What is 3D spatial annotation for robotics?
It labels how a robot can interact with objects through grasp sets, affordances, 6-DoF pose, functional parts and articulation.
How is this different from LiDAR annotation?
LiDAR perception annotation labels what is where; spatial manipulation annotation labels how to interact with it.
Why not one correct grasp?
There is no single correct grasp. Quality depends on coverage, exclusion, task conditioning and the end effector.
Can you verify automated annotations?
Yes. We review feasibility, constraints, hard cases, coverage and systematic generator failure patterns.
Is simulation enough?
Simulation is a filter, not proof. Hardware results are reported separately when access is in scope.
How do you handle symmetry?
Equivalent rotations are declared per object and handled explicitly in annotation and evaluation.
How is 6-DoF ground truth established?
Methods include markers, motion capture, CAD alignment, refinement and multi-view consistency.
Can you prepare articulated assets?
Yes, including joints, parts, collision meshes, physical properties and actuated URDF or MJCF validation.
How is quality measured?
Coverage overlap, exclusion agreement, task consistency, physical verification and geometric error where applicable.
How quickly can you start?
Verification can reach first delivery in about two weeks; from-scratch work usually reaches a cleared pilot in about three weeks, subject to scope.