How can I control the task sample?

[superggb]

In the imitation_learning example, the script will sample several demos as data to train. If I want to control the sample data, for example choosing a specified location to place the red ball to reach-target. Is there a convinient way to sample the demos I want?

[stepjam]

You can override tasks and add/edit behaviour. For example, in your case, you could do something like this:

class ReachTarget(Task):

    def init_episode(self, index: int) -> List[str]:
         desc = super(ReachTarget, self).init_episode(index)
         # explicit set the positions, e.g.
         self.target.set_pose(...)
         return desc

Then when you create the rlbench env, you can pass static_positions=True, which will disable automatic task placement.