I would like to have more control over espresso simulations and to explain this I will give an example.

I want to implement the "wonderful life" training policy in parallel. In order to do this, I need to create one simulation with N particles normally. However, I then need to create N copies of this environment, each with identical position and velocity initialisation, but, with one particle removed. The simplest thing in my mind would be the following:

simulation_runner = MyRunnerSimulation(...)

new_sim = deep copy(simulation_runner)
new_sim.remove_colloids(ids = [])

This may already be possible but I am not sure how to do such a thing.

What feature would you like to see added?
It would be really cool to be able to add tasks together from the front end API. Something like:

task = srl.tasks.RotateRod(...) + srl.tasks.FindLocation(...)

What feature would you like to see added?
I would like to see a clear vision cone observable added to SwarmRL as it seems there are a few variations floating around.

I am writing this here so that I can keep track of my progress on the two tasks. If you have any comments or recommendations, please also comment here so we can discuss it as it is implemented.

I am implementing two types of distributed strategies in SwarmRL, mainly focused on the RL.

• GeneticTrainer: In this gym, N RL model will be created and trained simultaneously. After M epochs, the one with the best read will be chosen, copied, and then more simulations run. This should avoid problems of having one or two bad runs and one good one as we will also pick the best.
• SharedGradientsTrainer: In the shared gradients gym, one RL model will run N simulations at the same time. During training updates, all gradients will be computed and then back-propagated through the network so that it can sample phase space faster.

Describe the bug
If I try to use the external force function of the espresso simulations I kill my kernel every time. I have tried it with the latest espresso but only from a jupyter environment not directly from the script.