[Question] canonical way to implement synchronous actions?

[tlpss]

I'm interested in using mujoco and dm_control to simulate a robot that can execute a trajectory for each step in the environment (e.g. for simulating a motion primitive with spatial actions).

In pybullet (or mujoco w/o the dm_control framework) in environment.step you would do IK, set the joint targets and call p.stepSimulation() untill the target pose is reached according to some criterion or just teleoport the joints to this state, collect the observations and return. But in dm_control, you cannot do this as the callbacks should be executed in time (so that other entities could also update their state based on the physics state or the action).

The number of physics steps required to execute this trajectory can vary quite a bit (nearby pose vs far away, or even teleporting the robot if it is not interacting with objects at that moment), so just running the worst case number of physics steps for each environment step seems very inefficient.

I'm wondering if this kind of synchronous (= "runtime-validated stop condition and hence variable number of physics steps") actions can be realised more efficiently in the dm_control framework, and what the canonical way is to do so.

One way I can think of, is an 'early termination' in the inner loop in the environment.step method, if the task signals this to the environment (the task itself can query its entities if needed), but that is not facilitated by the current Environment and I feel like I can't be the first person with this question so I wanted to hear if anyone knows a way to do this within the current framework?