JisuHann
commented
2 years ago 1. Multi-View Picking: Next-best-view Reaching for Improved Grasping in Clutter
Goal
- task: Visual grasp detection with active perception
- especially in clutter where many occlusions are present
- Difference with our problem formulation
- Ours: static camera position <-> Multi-view perception(Multiple fixed viewpoints)
Idea
- Active perception approach
- Use the act of reaching towards a grasp as a method of data collection
- Choose informative viewpoints(information-gain based approach, entropy) based directly on a distribution of grasp pose estimates in real time -> reduce uncertainty in the grasp pose caused by clutter and occlusions
- -> allow the system to be optimized for either raw grasp success rate or overall efficiency
- From A. Viewpoint Trajectory, Combine observations at time-steps along the view point
- From B. Visual Grasp Detection, Predict Grasp
- From C. Grid Map Representation, Update global Grid Map
- From D. MVP Controller
A. Viewpoint Trajectory
- Camera viewpoint p
- Viewpoint trajectory for grasping -> a set of K discrete viewpoints
B. Visual Grasp Detection: GG-CNN
: Real-time Generative Grasping Convolutional neural network
- input: depth image
- output: pixelwise visual grasp detection(Distribution of grasp estimation)
- (Quality, Angle=rotation around the vertical axis, Width=grasp width for a grasp at the position corresponding to each pixel)
C. Grid Map Representation
- represent the workspace of the robot as a 2D grid map (based on entropy)
D. MVP Controller
- Calculate the entropy of the grasp quality observations(in the quality observation) a single grid
- Approximate the expected information gain from an observation at a viewpoint
Results
- Success Rate
- Mean Time per Pick
- Mean Picks per Hour(MPPH): overall efficiency of the system: overall efficiency(average rate of successful picks per hour)
Manipulation Planning