Microsoft/malmo

Project Malmo is a platform for Artificial Intelligence experimentation and research built on top of Minecraft. We aim to inspire a new generation of research into challenging new problems presented by this unique environment.

Lennart

@biggzlar

@AndKram thanks will do.

Matthew Lee

@mattthelee

Hi. I've been running the find the goal v0 mission in Marlo. I keep finding that the environment does not give the assigned reward (0.5) back when an action finishes the game. It seems to depend on when i run the env.step afterwards. I've noticed a lag between taking an action and the position returned by info, is this related?

Matthew Lee

@mattthelee

nvm, found the solution is to put in a value for the step_sleep param to get Marlo to wait for Malmo

Shunta Saito

@mitmul

Hi, I'm using Malmo by launching it by docker run command introduced here:
https://github.com/Microsoft/malmo/blob/master/scripts/python-wheel/README.md#using-the-prebuilt-docker-image
and I want to communicate with the minecraft server from outside of the docker container using Marlo (https://github.com/crowdAI/marLo). I opened the port 10000 for the communication but the agent seems not to be able to communicate with the minecraft server in the docker container... Do I need to open another port? How can I know what ports should be opened to do it?

Andre Kramer

@AndKram

@mattthelee Rewards should be reported at mission end as they are sent with the mission end message but location and rewards during the mission are reported asynchronously. I think it would be an error to get any rewards after marlo env reports done == True but not an error to get a reward at some step after reaching a location.

Andre Kramer

@AndKram

@mitmul Malmo opens dynamic ports to talk back to the agent host from Minecraft (client) after the initial connection (here on port 10000) and on Windows docker's -net=host mode does not automatically open these unfortunately. Malmo (not marlo) allows you to specify the port number to use for the "command port" which is used to send commands during the mission though you would need to modify marlo to set that port so that it is not chosen dynamically. But there are other connections back to the agent host for video, rewards and observations which are harder to set up (in the mission XML) so it may be simpler to try to open a range for dynamic allocated ports. The port range is 10000 to 11000 so you could look to opening all these ports. Longer term, there is a new implementation of malmo for gym environments that only uses a single connection (https://github.com/Microsoft/malmo/tree/master/MalmoEnv).

Lennart

@biggzlar

Hi there. I have a question about the availability of object labels. This Microsoft/malmo#66 states that object labels have been integrated in 0.31.0 but I can’t find any documentation on that.
In this Microsoft/malmo#506 the poster is advised to use observationfromray, an XML tag. Can anyone tell me how to use these or alternatively how to use the apparently available object labels integrated in 0.31.0?

Lennart

@biggzlar

To add to my question, is there a method in place to translate world coordinates to screen coordinates?

Jewel James

@jeweljames

hey @AndKram - i have observed that after 123 back to back episodes, the agents run into
INFO:marlo.base_env_builder:Mission ended: MALMO_AGENT_NOT_RESPONDING
and then enter a loop of
enerating join tokens locally... INFO:marlo.base_env_builder:params.gameMode : Cannot force survival mode. INFO:marlo.base_env_builder:params.gameMode : Cannot force survival mode. INFO:marlo.base_env_builder:params.gameMode : Cannot force survival mode. INFO:marlo.base_env_builder:Waiting for mission to start... WARNING:marlo.base_env_builder:Error on attempting to start mission : Failed to find the server for this mission - you must start the agent that has role 0 first. WARNING:marlo.base_env_builder:Will attempt again after 3 seconds.

from which there is no recovery. in between it goes into other logs like
Pause for restarts .... Stopping ClientInfo: 127.0.0.1:10001:0 ... Stopping ClientInfo: 127.0.0.1:10000:0 ... Restart Malmo Minecraft clients for experiment:<experiment_id>

has this issue been observed before ?

Jumpei Hirono

@jhirono

Just sharing our malmo hands-on event pic. ;)

Andre Kramer

@AndKram

@jeweljames What OS are you on? There is a periodic restart of Minecraft as well as retries that should help if Minecraft becomes unstable or crashes. One issue on Windows is running in the same Minecraft directory. Copy the complete Minecraft dir (containing launchClient.bat) and run the second Minecraft client from that will help. If it happens again could you attach the logs (in run/logs - most likely just latest.log) and if Minecraft appears stuck then do a jstack command on the gradle process (you can find that using jps command) please? One other thing, some timeouts on multi-agent missions are over 90 seconds so it's best to wait ~3 mins to see if it recovers. Thanks, Andre.

@jhirono Great to see! Thanks for sharing.

Lennart

@biggzlar

Was trying to run launchClient.bat on Windows 10 and encountered this error. Any way to fix it?

Where:
Build file 'C:...\Desktop\malmo-master\Minecraft\build.gradle' line: 13

What went wrong:
An exception occurred applying plugin request [id: 'net.minecraftforge.gradle.forge', version: '2.0.1']
Failed to apply plugin [id 'net.minecraftforge.gradle.forge']
Unable to obtain url ([http://files.minecraftforge.net/maven/net/minecraftforge/forge/json]) with etag!

Jewel James

@jeweljames

@AndKram Thanks for answering.
Im observing this on ubuntu . im already running the minecraft instances from two folders.
im using the following params

                                    "client_pool": client_pool,
                                    "agent_names" : [
                                        "MarLo-Agent-0",
                                        "MarLo-Agent-1"],
                                     "step_sleep":0.01,
                                     "kill_clients_after_num_rounds":100,
                                     "kill_clients_retry":10
                                 }

this is reproducible as it happens around the same episode number each time. will attach the log.

Jewel James

@jeweljames

@AndKram - you can find the logs here - https://gist.github.com/jeweljames/788a8517ea76b02fa8a40970ee934398
i have attached the logs from the agents , the two minecraft instances(latest_m1.log and latest_m2.log) and two jstack logs.

ProQianXiao

@ProQianXiao

@AndKram @spMohanty Hello, are there any solutions that decides which mission is running?

Andre Kramer

@AndKram

@ProQianXiao My post from Nov 6th should provide a way to check which of the three games is being played. (Hope I understood the question correctly).

ProQianXiao

@ProQianXiao

@AndKram I tried that , but I met some errors:

Andre Kramer

@AndKram

@jeweljames Both Minecrafts seem to running a mission while one agent host (called from Python) is waiting on second to turn up and the second agent host is saying that all ports are in use. Which would indicate that previous missions are not finishing. Are you calling reset() before the step() function returns "done"? I've opened issue #784 on github malmo for this but have not seen this failure.

ProQianXiao

@ProQianXiao

I added the codes at the begining of the 'run_agent' function, is that right ?

Jewel James

@jeweljames

@AndKram im sometimes calling reset() before the step function returns "done". but only after calling close() on that env first.
is that a problem?

Andre Kramer

@AndKram

@jeweljames Better to not call close and re-use the env with reset(). It looks like there is a port leak when closing and the mission is still running? Let's take this to #784.

@ProQianXiao etree is not defined. I think we had this also in MarLo.: from lxml import etree.

ProQianXiao

@ProQianXiao

@AndKram Thanks and that helps. I also want to know whether we can use this method to dertermine the running misson when evaluation.

andrekramer

@andrekramer

@ProQianXiao (Andre on non-work account:) I think so but have not tried by submitting a test.

ProQianXiao

@ProQianXiao

@andrekramer Thank you, and I will have a try.

iZuno-K

@iZuno-K

@andrekramer @AndKram I got errors at auto-evaluation phase:
“Unable to orchestrate evaluation. Core services failed to start…”
and after 1 or 2 episodes:
“Unable to orchestrate submission, please contact Administrators.”
The submitted programs have succeeded the test_submission_locally.sh without errors.
(Issue url:
https://gitlab.crowdai.org/karino_isi/marlo2018/issues/8
https://gitlab.crowdai.org/karino_isi/marlo2018/issues/5
)

Could you check the cause and teach me how to fix it ?
My CrowdAI account is karino_isi.

Lennart

@biggzlar

One more question: suppose I have several different models - is there a way to access the gametype during evaluation?

SP Mohanty

@spMohanty

@iZuno-K : We are investigating the error. and Will get back to you guys soon.

SP Mohanty

@spMohanty

I just tried to make a new submission with the starter kit, and it went through completely fine.

So theres a chance that this might be because of some issue in the actual submitted code.

iZuno-K

@iZuno-K

@spMohanty Thank you for your investigating.

I’ve succeeded the local test based on crowdAI’s instruction (https://github.com/crowdAI/marlo-multi-agent-starter-kit). I tested 1 game with 2 agents and 3 different games with 6 agents. Both tests succeed without errors.

How can I check my submitted code have any issue?

SP Mohanty

@spMohanty

@iZuno-K : Do you test it for multiple episodes locally ?

iZuno-K

@iZuno-K

@spMohanty I tested multiple episodes today. My code cause an error RuntimeError: Attempted to use a closed Session. The reason is I use tf.Session().close() at the end of the run_episode().
Then I modified run.py by moving the tf.Session().close() line to the end of the main function. After that I tested 15 episodes local test and succeeded.
I submitted the modified code but receive Unable to orchestrate evaluation. Core services failed to start....

ProQianXiao

@ProQianXiao

@spMohanty @AndKram I submitted my code but I got the following error ,

it seems that episode-1 and episode-2 are successful, but I do not know why the later episodes failed.

ProQianXiao

@ProQianXiao

@spMohanty @AndKram How to solve the problem "Unable to orchestrate evaluation. Core services failed to start...."? Is it a evaluation environment error or code error?

iZuno-K

@iZuno-K

@ProQianXiao @spMohanty @AndKram I solved the problem by removing apt.txt from my repository.

I removed these dependencies and then the evaluator works.
libopenmpi-dev, openmpi-bin (for mpi4py)
libgtk2.0-dev, libglib2.0 (for opencv)

ProQianXiao

@ProQianXiao

@iZuno-K Thanks for your reply, did it work for you ? I checked my environment.yml, in fact there are no these dependencies.

This is mine

iZuno-K

@iZuno-K

@ProQianXiao @spMohanty @AndKram Sorry, that is not the cause... I tried a new submission now with the same dependency with only modified trained parameters. But it failed...

ProQianXiao

@ProQianXiao

I always met this error, did you have same error?

Where communities thrive

People

Repo info

Activity