Adds tensorflow support and async TF ros controller.

bstadie commented 8 years ago

This pull request adds tensorflow support. In particular, it allows badmm examples to utilize a tensorflow backend. The following additions are made:

PolicyOptTf class for optimizing policy in tensorflow.
TfPolicy class for running tensorlow policies forward.
Auxiliary classes TfSolver and TfMap which allow reproducible specification of tensorflow models.
ROS agent receives a new tensorflow controller, which allows for async control of the robot, with observations being published in the ROS CPP agent, received by the python agent, and then run forward though tensorflow in python. The resulting actions are published back to ROS where they are received by the CPP agent and used to directly set joint torques.
In addition to ROS, TF policy optimization is supported by Box2D and MuJoCo agents (I have tested them on my machine).
Add a tests folder and relevant tests for the newly added code.

cbfinn commented 8 years ago

This looks great Bradley! I'm about halfway through looking at it. I'll resume tonight.

cbfinn commented 8 years ago

Other than the two code comments above, this is good to go. Feel free to merge after fixing.

bstadie commented 8 years ago

I can not merge, as I do not have write access to this repo.

cbfinn commented 8 years ago

Just gave you write access for the future. I'll merge this PR now.

cbfinn / gps