Auto-Labeling for 6-DoF Grasp-Pose Detection
An RGB-D auto-labeling pipeline that generates 6-DoF grasp-pose annotations for robotic pick-and-place in unseen domestic environments.
Project Snapshot
| Signal | Details |
|---|---|
| Timeline | Mar 2023 to May 2023 |
| Organization | XPENG Robotics |
| Focus | Implemented the Auto-labeling pipeline to Generate Annotations for Grasp Pose Detection |
| Stack | Custom implementation and project-specific tooling |
Core Contribution
- Built a grasp pose annotation workflow for unseen objects and novel scenes.
- Used antipodal grasp sampling to generate candidate grasps.
- Implemented a new method following the method mentioned in GraspNet-1Billion dataset paper (GraspNet-1Billion code is not public, we just re-implemented method following the details from the paper).
Pipeline
- Step 1: Generate Grasp Poses on Object Mesh Model
- Step 2: Generate 6D Pose of the Object (Can use ICP or explore other methods)
- We manually label the 6D Pose of the object, so don’t need a algorithm to generate 6D Object Pose
- Step 3: Projection of generated grasp poses on the scene using 6D Object pose of each object
- Step 4: Using camera poses to transform generated grasps into each view
- Step 5: Filtering grasps using collision detection
- Collision detection is conducted to avoid the collision between grasps and background or other object