Mujoco 1.0 #733
Mujoco 1.0 #733
Conversation
|
Thank you, very useful! I'm using your code with the latest MuJoCo. |
| high = bounds[:, 1] | ||
| self.action_space = spaces.Box(low, high) | ||
| bounds = self.model.actuator_ctrlrange | ||
| # I'm not sure why bounds is at least sometimes None ... bug? |
There was a problem hiding this comment.
@adamlerer / @yayitsamyzhang could you share examples of when this happens? Possibly a bug.
There was a problem hiding this comment.
Unfortunately I've lost track of when this happened. You could try setting it to None and then run an environment and see what happens!
There was a problem hiding this comment.
It's constructed by generated code (here in the most recent version: https://github.com/openai/mujoco-py/blob/master/mujoco_py/generated/wrappers.pxi#L1285).
Wrapper is defined here: https://github.com/openai/mujoco-py/blob/master/mujoco_py/generated/wrappers.pxi#L3824
So it looks like we set it to None instead of a numpy array with 0 as one of the dimensions (which would just be empty anyways).
This would be true for models without any actuators (so all the ctrl related arrays are size 0).
I'm alright with this handling -- maybe update the comment to say that it's None when there are no actuators.
If you wanted to, you could check sim.model.nu == 0, which is the same thing (no actuators in the model).
There was a problem hiding this comment.
This looks great! I'm looking forward to trying it out.
Given that this changes the underlying simulator, the environment name versions Humanoid-v1, etc need to be bumped, to show that the environment has changed.
If any of y'all are able to do it, it'd be good to try training policies for all of the old environments, and verify they work well on the new environments and vice versa.
| self.model.forward() | ||
| sim_state = self.sim.get_state() | ||
| sim_state.qpos[:] = qpos | ||
| sim_state.qvel[:] = qvel |
There was a problem hiding this comment.
I'm pretty sure what we want to do here is
self.data.qpos[:] = qpos
self.data.qvel[:] = qvel
The state is a separate thing, and if you set those parameters on it they'll just get thrown away.
|
This is addressed in #834. |
Some people have been asking for mujoco-1.0 (1.5.0.x) integration with gym (e.g. for headless rendering). This got it working for us.