1. Paper Presentation --- Grasping a novel object InInstructor: Student Name: Major: ID: Date: Marius C.Silaghi Sida Du M.E. 901723499 April 16 2014

2. Learning Grasp Strategies with Partial Shape Information  The main problem: Grasping a novel object in a cluttered environments  Why need a new algorithm: Old method couldn’t perceive well with complex condition  The strategy of the new algorithm: Algorithm dealing with features from two kinds of sensors  Results of grasping with that: Better than ever

3. Introduction  The background:  Mason and Salisbury 1985; Bicchi and Kumar 2000; Pollard 2004  A earlier work on Saxena et al.

4. Introduction  The situation:  cluttered environments  Two kinds of senors  Robot with fingers  Given model of object

5. Introduction  Description of Robots:  STAIR 1 uses a 5-dof harmonic arm with a parallel plate gripper, and STAIR 2 uses a 7-dof arm with a three-fingered 4-dof hand.  The robot’s vision system consists of a stereo camera and a SwissRanger camera.

6. Grasping Strategy  Definition of grasp  The full goal configuration of the arm/fingers is given by α.  The full goal configuration of the arm/fingers that is required to grasp an object will be referred as a grasp.

7. Grasping Strategy  Obtain the data from sensors  We will use both the point-cloud R and the image I taken of a scene to infer a goal configuration of the arm/fingers.

8. Grasping Strategy  Features of grasp  There are a variety of features that should be considered: minimizing slippage, sufficient contacting with the object, distance to obstacles, and distance between the center of the object and the grasping point.

9. Grasping Strategy Estimating features vector  Presence / Contact:  For a given finger configuration, some part of an object should be inside the volume enclosed by the fingers.

10. Grasping Strategy Estimate features vector  Symmetry / Center of Mass:  Even if many points are enclosed by the hand, their distribution is also important.

11. Grasping Strategy Estimate features vector  Local Planarity / Force-Closure:  The direction of force applied by fingers should be in a flat or Intersect

12. Grasping Strategy Capture such properties  Starting with calculating the principal directions of a 3-d point cloud centered at the point in the question.  This gives three orthonormal component directions ui, with u1 being the component with largest variance, followed by u2 and u3. Let σi be the corresponding variances.

13. Grasping Strategy Capture such properties  For fj as the finger direction (j = 1 for parallel gripper, and j = 1, 2 for three-fingered hand)  Ideally, the finger direction should be orthogonal to large variance directions and parallel to the small variance ones.  Calculate the following features:  (a) Directional similarity, sij =  (b) Difference from ideal,

14. Grasping Strategy  Probabilistic Model  Use a similar classifier that computes a set of image features and predict the probability P(y = 1|α, I) in [0, 1] of each point in the image being a candidate grasping point.  Use a second classifier that, given a configuration and a point-cloud R, predicts the probability P(y|α,R) that the grasp will succeed.  Combine these two classifiers to estimate the probability P(y|,R, I) of a grasp succeeding.

15. Grasping Strategy  The inference probability  the P(y|a, I) is the 2-d image classifier term similar to the one in Saxena et al. (2006a).  The P(y|α,R) is computed with  After that,

16. Grasping Strategy  The inference probability

17. Grasping Strategy  Noting that a configuration has a very small chance of being the optimal configuration if either one of the two terms is very small.  Given one such 3-d location, finding a full configuration for it now requires solving only an n − 3 dimensional problem.  It was sufficient to consider only a few locations of the fingers, which further reduces the search space

18. Experiments  With the new approach :  Testing this algorithm on two robots, on a variety of objects of shapes very different from ones in the training set, including a ski boot, a coil of wire, a game controller, and so on.

19. Experiments  Grasping single novel objects  We considered several objects from 13 novel object classes in a total of 150 experiments. These object classes varied greatly in shape, size, and appearance, and are very different from the plates, bowls, and rectangular blocks used in the training set.

20. Experiments

21.  Grasping in cluttered scenarios  In a total of 40 experiments, the success rate was 75%, and this robot is the first one to be able to automatically grasp objects, of types never seen before, placed randomly in such heavily cluttered environments.

22. Experiments

23. Experiments movies http://stair.stanfo rd.edu/movies/a aai_3d_grasping _very_clutter.wm v

24. Experiments movies http://stair.stanfo rd.edu/movies/a aai_3d_grasping _kitchen_sequen ce.wmv

25. Thank you for attention! ANY QUESTION