Registered version details for Search and Rescue environment?

[Chulabhaya]

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?

[sash-a]

Hi @Chulabhaya 👋

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

[zombie-einstein]

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.

[zombie-einstein]

There's some notes in the original PR for the environment, may help. Though apologies they are not super detailed!

[Chulabhaya]

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.

[sash-a]

All the props for that goes to @zombie-einstein! I just reviewed it 😄

But totally agreed, we had a similar issue when writing this paper we could only find MPE and MaBrax

[zombie-einstein]

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!

[sash-a]

Whenever you've got the time, contributions are greatly appreciated!

[Chulabhaya]

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.

[Chulabhaya]

@zombie-einstein Hi there, happy holidays! I was just reaching out because I'm at the stage of wanting to benchmark a few variations of Search and Rescue for a paper I'm working on, and I was wondering, like the Connector environment, is it possible to have a wrapper for this environment such that the observation is a flattened vector? Or would that not make sense for this environment in your opinion?

[zombie-einstein]

Hi @Chulabhaya, you mean to effectively treat as a single agent environment?

Yeah it's definitely possible, I think I did something similar to this when I was testing the environment, for example you could use the positions of the agents as values in the observation. Or could have a very large concatenation of the individual observations.

I guess the only potential drawback is that you then have an observation size dependent on the number of agents in the environment.

[Chulabhaya]

Hi @Chulabhaya, you mean to effectively treat as a single agent environment?

Yeah it's definitely possible, I think I did something similar to this when I was testing the environment, for example you could use the positions of the agents as values in the observation. Or could have a very large concatenation of the individual observations.

I guess the only potential drawback is that you then have an observation size dependent on the number of agents in the environment.

Hmmm, so what I was envisioning was something like essentially instead of a 2D grid as an observation for each agent, the agent instead gets a flattened version of that grid that's just some long 1D feature vector. The environment would still be multi-agent since each agent would have a unique flattened obs vector. This would be to just avoid needing to mess with needing convolutional layers for pre-processing the grid and just being able to pass the input directly into a simple FC layer. Does that seem reasonable? I'd imagine then the global state is just a concatenation of all the flattened obs of all agents.

[zombie-einstein]

Oh right yeah this should work fine. Each row (of each agents) observation is just a view on different element types in the environment, so flattening it should be fine.

I guess it may be slightly suboptimal given that there is some 2d spatial information in the observations, but apart from that it's perfectly reasonable.

[Chulabhaya]

Oh right yeah this should work fine. Each row (of each agents) observation is just a view on different element types in the environment, so flattening it should be fine.

I guess it may be slightly suboptimal given that there is some 2d spatial information in the observations, but apart from that it's perfectly reasonable.

On second thought, I'd actually like to just keep the obs as-intended to avoid any unintended consequences and probably just use a CNN. When you did your training, did you happen to set your Mava network to mava/configs/network/rcnn.yaml?

[Chulabhaya]

Hmmm, so I seem to be having issues getting this to work in general. I went through this fork (https://github.com/zombie-einstein/Mava) and copied over the configs and environment wrappers. When actually running the environment however I don't see a 2D observation, rather the agent view gives me a 1D vector. I am unsure whether I am doing something wrong, or if there is some compatibility issue between the SAR wrapper in the fork vs. the latest version of Mava. Any ideas?

My environment config is the following:

# ---Environment Configs---
defaults:
  - _self_
  - scenario: sar-80tg-8ag

env_name: SearchAndRescue  # Used for logging purposes.

# Defines the metric that will be used to evaluate the performance of the agent.
# This metric is returned at the end of an experiment and can be used for hyperparameter tuning.
eval_metric: episode_return

# Whether the environment observations encode implicit agent IDs. If True, the AgentID wrapper is not used.
# This should not be changed.
implicit_agent_id: False
# Whether or not to log the winrate of this environment. This should not be changed as not all
# environments have a winrate metric.
log_win_rate: False

kwargs:
  {}

And my scenario config is the following:

name: SearchAndRescue-v0
task_name: sar-80tg-8ag

task_config:
  num_searchers: 8
  num_targets: 80

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