When turn-taking breaks down
Synchronization between agents in worlds gets trickier when the world is operating in real time, as with robots and other physical systems. In non real-time RL, the world can wait for the agent to plan its next action. It is a turn-taking scenario where the agent waits for the world to make its next move and vice versa. But for physical hardware the world doesn't wait for the agent to compute. The instant it reports its sensor readings to the agent, they are already obsolete. The world has moved on in its machinations. Time and tide wait for no agent.
If the world moves slowly compared to the speed at which the agent computes, this mismatch can be mostly ignored. But if computations are intensive, as with some vision applications and nearly all transformer-based agents, or compute power is limited, as with many mobile robots, then this delay can become considerable. It can render modern algorithms infeasible.
This problem is described in this 2019 paper as well as here and here. The terminology used to describe it varies, but if you want to explore the literature, the search terms "stochastic delay" and "non-integer delay" will get you started.
A bad solution: Waiting around for the agent
One workaround would be to have the world sit idle and take no action for a short period to enable agent to run itself, but this would require timing in the world to be tuned to the needs of the agent, which is dependent on the algorithm, the hardware, and other things that the world should not have to know about.
An OK solution: Predict the near future
These authors propose a solution where action selection is allowed a one time step delay, and the previous action is included in the state. This allows the agent to indirectly predict what the state of the world will be once its action command is finally executed. It demonstrates the philosophy of "Skate to where the puck is going".
One interesting case study here is the human brain moving the musculoskeletal system. The ability to predict the state of the system a short time into the future appears to be a critical component of its robust performance. (Although it does seem to be very sensitive to changes in the length of that delay.)
This predictive approach has the benefit of being agnostic to the system being controlled and the computation hardware being used to select the actions. It obtains this benefit by making the assumption that one time step is roughly the amount of time needed to choose an action. If the time is much shorter than that, say one tenth of a time step, this approach adds a gratuitous delay of 9/10 of a time step. And in cases where action selection might take several time steps, for example where more expensive planning or slower retrieval processes are involved, it would still fail to account for that.
A better solution (IMO): React as soon as the agent's command is available
A better way to work around this is to take advantage of the fact that robots typically have a fast internal clock, and to have them try to read a new action at every step (or few steps) of this hardware clock, rather than at every agent step. This will ensure that as soon as the new action is available, it will get incorporated into the world.
In practice this will introduce a small delay of several hardware steps. I don’t know of any way to avoid this. This is an important feature of real time reinforcement learning that gets lost in simulations. Any real time controller or physical hardware has to deal with this delay.