instadeepai
Registered version details for Search and Rescue environment?
Hello! I'm very interested in using the Search and Rescue environment for a project, but unfortunately, unlike the other Jumanji environments, there doesn't appear to be a registered version based on the documentation: https://instadeepai.github.io/jumanji/environments/search_and_rescue/
What settings represent a good baseline, similar to what's shown in the documentation page animation with 4 agents?
Hi @Chulabhaya đ
Glad you're interested in search and rescue. Thanks for catching the issue in the docs, we should add to the documentation.
As with most jumanji envs there is only 1 registered version: SearchAndRescue-v0. However it is easy to create scenarios by modifying the generator and the env with arguments to jumanji.make. For example:
jumanji.environments.swarms.search_and_rescue.generator import RandomGenerator
env = jumanji.make("SearchAndRescue-v0", generator=RandomGenerator(num_targets=100, num_searchers=10), searcher_max_speed=0.02)
There are other arguments you can change, you can see them here.
By default num_targets is 40 and num_searchers is 2. I am fairly certain this is what we used when testing the environment, but maybe @zombie-einstein could confirm? I'd encourage you to start with the default and then play around with the options as you see fit
Hey @sash-a and @Chulabhaya,
Yeah the defaults are pretty good, when I was testing with Mava I was using these settings (with the rest of the defaults used):
name: SearchAndRescue-v0
task_name: default
task_config:
num_searchers: 4
num_targets: 40
env_kwargs:
target_contact_range: 0.02
searcher_max_rotate: 0.25
searcher_max_accelerate: 0.005
searcher_min_speed: 0.005
searcher_max_speed: 0.02
time_limit: 400
which seemed to generate a reasonably challenging, but still manageable environment.
There's obviously a lot of scope to vary parameters, and they can be quite big changes in the difficulty and the nature of the task, especially around the density and visibility of the targets.
There's some notes in the original PR for the environment, may help. Though apologies they are not super detailed!
Thank you guys @zombie-einstein @sash-a , this was extremely helpful! Also huge thanks for adding this environment, there's a distinct lack of good JAX MARL environments that focus on continuous actions so it's really nice to see environments like this get created.
All the props for that goes to @zombie-einstein! I just reviewed it đ
But totally agreed, we had a similar issue when writing this paper we could only find MPE and MaBrax
Thanks @Chulabhaya glad you are using it. I should have also said I forked Mava when I was testing, this may also help you out.
I am planning to implement some more environment along this theme. I've had some other commitments taking up my time but need to get back on this @sash-a!
Thank you both for the comments! I appreciate the Mava recommendation, originally what I was trying to do was get some Jumanji environments working with JaxMARL's algorithm implementations, but I think Mava algorithms have better integration across both environment frameworks so I'll start there.