1

I trying to use DDPG augmented with Hindsight Experience Replay (HER) on pybullet's KukaGymEnv.

To formulate the feature vector for the goal state, I need to know what the features of the state of the environment represent. To be precise, a typical state vector of KukaGymEnv is an object of the numpy.ndarray class with a shape of (9,).

What do each of these 8 elements represent, and how can I formulate the goal state vector for this environment? I tried going through the source code of the KukaGymEnv, but was unable to understand anything useful.

nbro
  • 39,006
  • 12
  • 98
  • 176
Vedant Shah
  • 105
  • 1
  • 7

1 Answers1

1

Here's an incomplete answer, but it may help.

Your state is read by the function getExtendedObservation(). This function makes two things : it calls the function getObservation() from this source code, gets a state, and extend this state with three components :

relative x,y position and euler angle of block in gripper space

But what are the 5 first components returned by getObservation()? From what I read, there are positions, then euler angles describing the orientation. But that would make 6 + 3 = 9 features, so there is either only 2 positions, or only 2 euler angles. You may know kuka better than me and know the answer of this one :).

So, to sum up :

state = [X, Y, (Z, ) , Alpha, Gamma, (Beta, ), gripX, gripY, gripAlpha]

(Either Z or Beta is absent)

16Aghnar
  • 591
  • 2
  • 10
  • 1
    There's a mistake in the question. The state space for the environment is actually a 9 dimensional vector, in which case, your answer is correct. I ll rectify the typo right away. Thanks for the help! – Vedant Shah Aug 27 '20 at 10:38