Grasping example - #371
Conversation
prabhatnagarajan
left a comment
prabhatnagarajan
left a comment
There was a problem hiding this comment.
Looks good. Take a look at my comments before merging.
|
|
||
| # Watch how the learned agent performs | ||
|
|
||
| python examples/pybullet/train_dqn_batch_grasping.py --demo --render --load <path to agent> |
There was a problem hiding this comment.
I think it would be helpful if we were more specific about what is. Is it the same path as the outdir flag that is passed in to the agent during training?
There was a problem hiding this comment.
I clarified this in README.
|
|
||
| def __call__(self, x): | ||
| image, steps = x | ||
| h = self.image2hidden(image) * F.sigmoid(self.embed(steps)) |
There was a problem hiding this comment.
Do we really need to divide this into two steps? Can the neural network not be implemented in one shot?
There was a problem hiding this comment.
It is possible to broadcast steps, concatenate it to image as an additional channel, and apply convolutions. The current implementation is simpler in a sense that it does not require broadcasting. I don't know which works better in practice.
2 participants
Merge #366 first.This PR adds an example of learning to grasp using a Bullet-based env.
Updated READMEs:
TODOs: