Research Topic
A robotic pick-and-place system that is capable of grasping and recognizing both known and novel objects in cluttered environments.
这里的novel object是指不需要task-specific training data for novel objects
Contribution
文章的贡献就是 提出了两个framework
- object-agnostic grasping framework: visual observations
→
\rightarrow
→ actions
得到dense pixel-wise probability maps of the affordances for four different grasping primitive actions.
(affordance 和 probability of picking success差不多意思) - cross domain image classification framework
match observed images to product images,起到object recognition的作用
Related Work
有两大类的方法:Recognition followed by Model-based Grasping
-
classical solutions for robotic picking——standard two-step approach:
- recognition and pose estimation
- model-based grasp planning
这些方法都需要3D models of the objects during test time, and/or training data with the physical objects themselves.
因此都很难扩展到novel objects.这篇文章提出的grasping framework其实没有object identity这步,因此自然也就object-agnostic了。
-
Recognition in parallel with Object-Agnostic Grasping
这类方法就是exploit local features without object identity了
they still need some form of in-place object recognition in order to associate grasp proposals with object identities. -
Active Perception
exploiting control strategies for acquiring data to improve perception.
这篇文章也是受这种思想启发,follow a g r a s p − f i r s t − t h e n − r e c o g n i z e grasp-first-then-recognize grasp−first−then−recognize work flow.
System Overview
overall approach
- uses FCNs to infer the pixel-wise affordances of four different grasping primitive action
- 选择affordance最高的primitive action, pick up the object, isolated it from the clutter, hold it up in front of the camera
- recognizes its category, and place it in the appropriate bin
system setup
- 6 DOF ABB IRB1600id robot arm
- 自己设计的末端装置,既能抓也能吸
- RealSense SR300 RGB-D cameras