Skip to content

Data and input variations

Stephanie Hughes edited this page Jul 28, 2021 · 6 revisions

Shapes

Shape of the object. All possible shapes options: Cube, Cylinder, Cube45, Vase1, Vase2, Cone1, Cone2, Bottle, Bowl, TBottle, RBowl

Sizes

Sizes of the object; determined by the hand-span. All size options: Small (S), Medium (M), Big (B

Orientations

The hand orientations are composed of normal (0 deg. rotation), slanted (45 deg. rotation), and top (90 deg. rotation) to adapt the grasping strategy to handle diverse hand positions as shown in the figure below.

Hand Orientation Variation (HOV)

Hand Orientation Variation (HOV) is the position of the hand given an additional rotation (ranging from +/- 5 degrees), w.r.t. the palm, from the original hand orientation (normal (0 deg), rotated (68 deg), top (90 deg)). (See HOV outlined in the figure above)

HOV hand-object pose dataset generation:

  1. Move the object within the graspable region of the hand
  2. Rotate the hand relative to the object (±5 degrees)
  3. Check each location to ensure no collision between the hand and object

Input Variations

The input variation defines a set of objects (determined by size/shape) and hand orientation (normal, rotated, top) that we use to set the starting conditions for the controllers and policy.

Baseline: Base input of a medium cube and normal hand orientation

Baseline + HOV: Same input shape and hand orientation as the baseline, with the addition of variation in the hand orientation for the hand and object starting position coordinates.

Object Size Variation + HOV: Input objects vary in size (small, medium, big) and sample from hand orientation variation coordinates

Object Shape Variation + HOV: Input objects vary in shape (cube, cylinder, vase) and sample from hand orientation variation coordinates

Orientation Variation + HOV: Input objects vary in hand orientation (normal, side, top) and sample from hand orientation variation coordinates

Hand Orientation Variation (HOV): Experiments with HOV sample the initial hand orientation from the HOV coordinate list. These coordinates are sampled by both the constant-speed controller to generate expert replay buffer experience and by the policy during training.