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.
@EC2EZ4RD @tdjogi010 Mohnaty is travelling so there may be some delay on MalmoAdm and I'll leave it to him to say more about the dates but I would submit early and often.
@AndKram - Can you please expand on this note
NOTE: If instead of a dynamic loop, you hard code the run_agent
function calls, then the evaluation of your code will fail in case
of a tournament, where multiple submissions can control different agents
in the same game.
how does this work in the tournament format?
@jeweljames I think it's saying that you should not assume two join tokens (one for each of the two agents in the game) but that you could be given just one token if the other agent in the game is some other submission's. As long as you stick with the general outline (i.e. the loop here in run.py) you should be fine.
@AndKram @spMohanty Hello, I was wondering if the rule presented in this paragraph; "Your team lead must notify Organizer before the end of the Entry Period (Nov 12, 2018) by sending an email to MalmoAdm@microsoft.com, indicating your crowdAI team name and all every individual team members. Only emails actually received by Organizer before the end of the Entry Period will be considered." is still in effect since the entry period has changed to 31st december? That is do you have to have a complete team and email it latest today to be able to compete for prizes in this competition?
@EC2EZ4RD The multi-agent environments are listed at the end of: https://marlo.readthedocs.io/en/latest/available_envs.html and Build Battle's main (https://marlo.readthedocs.io/en/latest/marlo.envs.BuildbattleTrain1.main.html#module-marlo.envs.BuildbattleTrain1.main) gives the details. Basically you have to copy the box structure before you.
@EC2EZ4RD @tdjogi010 @denissv : Sorry for the delay in the response from MalmoAdmn . All your emails have been received, and I will respond back with an acknowledgement right away.
And @dennissv : You can still register your team after November 12th, and I will have that line changed after checking internally with all the co-organizers.
And @dennissv : You can still register your team after November 12th, and I will have that line changed after checking internally with all the co-organizers.
Or, better, just ask the person responsible from Microsoft, to respond to you here, and make the changes accordingly.
@spMohanty I was trying to setup multi-agent environment but I am getting this error:
gym.error.UnregisteredEnv: No registered env with id: MarLo-BuildbattleTrain1-v0. After following the instructions in the marlo-single-agent-starter-kit (same as marlo-multi-agent-starter-kit, I realized, I must also install the repository https://github.com/crowdAI/marLo using python setup.py install. The repository marLo does contain the environment, but the error persists after the installation. How can I fix this?
@sushrutbhalla Possibly you are still picking up an old version of marlo? One common problem is different versions of python (should you have done python3 setup.py install?) or not activating the right anaconda environment (at install or on running the sample?). You can do "import marlo.envs; help(marlo.envs)" in python3 and see both the available marlo envs and (at end) the install path which may give you a comparison against what you installed. By the way, you should be able to do "pip3 uninstall marlo" and then "pip3 install marlo" to get the latest version with the new multi-user envs.
@freeone3000 There are more versions of the native wheel on test.pypi.org. Use pip3 install --index-url https://test.pypi.org/simple/ malmo. The anaconda install should find one if it's available for your OS.
I'm installing Malmo on OSX using pip and can launch minecraft however when I try to run the tutorials from terminal I get this error:
Traceback (most recent call last):
File "tutorial_1.py", line 24, in <module>
import MalmoPython
ImportError: dlopen(/Users/programming/Programming/Malmo/MalmoPlatform/Malmo/samples/Python_examples/MalmoPython.so, 2): Symbol not found: __ZNK5boost16re_detail_10670031cpp_regex_traits_implementationIcE17transform_primaryEPKcS4_
I would suggest best to use the anaconda env we released to avoid version incompatibilities
When I run the Multi Agent Example, this error occur:
File "/anaconda3/envs/marlo/lib/python3.6/site-packages/marlo/base_env_builder.py", line 577, in transform_mission_xml
vp = etree.Element(ns + "VideoProducer")
NameError: name 'etree' is not defined. I can run the single agent example though.
Also, when I submit the multi agent starter kit unchanged, it says evaluation. Is there anything I need to change besides the username in json?
*it says evaluation failed
@danger-mon Do you have boost installed? Malmo needs a compatible version of boost which you can get with brew: https://github.com/Microsoft/malmo/blob/master/doc/install_macosx.md
Hi, did you test using multiprocessing to run the code. I met a problem and hope to get some help.
I first used ctx = multiprocessing.get_context('fork') to create a multiprocessing context.
Then in a for .. in ...loop, I created process = ctx.Process(target=runner, args=(join_tokens, i, shared_buffer)), and in runner, I used join_tokens and i to init role i's env to run. Multi agent ran well in the first rollout. But when they finished the first rollout and then got to reset(), then the first role's minecraft worked well, waiting for the second to join. But the second role's minecraft crashed, with logging said:
[15:24:53] [Client thread/INFO]: [STDOUT]: CLIENT request state: MISSION_ENDED
[15:24:53] [Client thread/INFO]: [STDOUT]: CLIENT enter state: MISSION_ENDED
[15:24:53] [Client thread/INFO]: [STDOUT]: Sending mission ended message to localhost:10638.
[15:24:53] [Client thread/INFO]: [STDOUT]: CLIENT request state: WAITING_FOR_SERVER_MISSION_END
[15:24:53] [Client thread/INFO]: [STDOUT]: CLIENT enter state: WAITING_FOR_SERVER_MISSION_END
[15:24:53] [Client thread/INFO]: [STDOUT]: CLIENT request state: DORMANT
[15:24:53] [Client thread/INFO]: [STDOUT]: CLIENT enter state: DORMANT
[15:24:53] [Client thread/INFO]: [STDOUT]: JETTISONING 0 COMMANDS
[15:24:54] [Thread-22/INFO]: [STDOUT]: Received from localhost:
[15:24:54] [Thread-22/INFO]: [STDOUT]: MALMO_REQUEST_CLIENT:0.36.0:20000:954b2199-9382-4296-accd-1e2e8878582b
[15:24:54] [Thread-22/INFO]: [STDOUT]: ==== RES: - 6374
[15:24:54] [Thread-22/INFO]: [STDOUT]: REPLYING WITH: MALMOOK
[15:24:55] [Client thread/INFO] [FML]: Applying holder lookups
[15:24:55] [Client thread/INFO] [FML]: Holder lookups applied
[15:24:56] [Client thread/FATAL]: Unreported exception thrown!
java.lang.StackOverflowError
at net.minecraft.client.gui.FontRenderer.getFormatFromString(FontRenderer.java:984) ~[FontRenderer.class:?]
at net.minecraft.client.gui.FontRenderer.wrapFormattedStringToWidth(FontRenderer.java:887) ~[FontRenderer.class:?]
at net.minecraft.client.gui.FontRenderer.wrapFormattedStringToWidth(FontRenderer.java:888) ~[FontRenderer.class:?]
at net.minecraft.client.gui.FontRenderer.wrapFormattedStringToWidth(FontRenderer.java:888) ~[FontRenderer.class:?]I am always running my code in Ubuntu and the default test code works well. It uses marlo.threaded wrapper for multi-agent scenario. I also can train agent using it. But it is more hard to share some variables between threads than between processes.
Hi Guys, is it possible for send discrete commands to env returned by marlo.init in FindTheGoal? I can send discrete commands(e.g. 'movenorth 1', 'movewest 1') using agent_host.sendCommand(). This agent_host is returned by MalmoPython.AgentHost(). However when I use the env(which seems like a warpper to OpenAI gym), I can use env.step() method for continuous actions (actions from env.action_space). I can see a send_command() method too in env, but when I send discrete commands(e.g. 'movenorth 1', 'movewest 1'), the agent does nothing.
Hello folks,
We will do Minecraft x Reinforcement learning hands on using malmo/ChainerRL. https://dllab.connpass.com/event/108036/ Sorry it is in Japanese but demand is huge 100 registrations within hours.
- We will prepare Azure GPU env(K80 x 1) for each participants.
- https://github.com/mitmul/marlo-handson is the hands-on contents v0.1.
we have several challenges;
- Initially we try to use docker, but it doesn't work well since we cannot establish communication between script and malmo even we make ports open for host and communication
- minecraft screeon on x2go is hard to see. are there any good ways to show trained results?
- It is great if we can prepare other scenarios and multi agent training for hands-on. do you have any documents and scripts?
thanks for your help.
@pankajsaini0803 FindTheGoal has a mission xml (marlo/env/FindTheGoal/templates/mission.xml) that has ContinuousMovementCommands in the AgentHandlers section so the agent reacts only to continuous commands. You can of course change this (DiscreteMovementCommands) for experimentation but we had it as continuous only for the first round of the MarLo competition. The second round games are discrete but are also turn based in that each agent is given alternative turns to send commands.
@jhirono Sound like a great use of Marlo. On Windows the "host" networking in docker does not forward all ports so the Malmo agent can't communicate over dynamic ports which are opened after the initial connection. It should be possible to get around this by specifying ports in the XML or listing a range of dynamic ports using the -p option to docker run. A simpler new implementation of the Open AI gym environment uses one (or two if multi-agent) connections which works more simply with docker (https://github.com/Microsoft/malmo/tree/master/MalmoEnv). But the simplest approach may be to run everything (AgentHost and Minecraft) in the same docker image. You can use the docker image described on https://github.com/Microsoft/malmo/blob/master/scripts/python-wheel/README.md to remote the Minecraft display over VNC (into a Web browser, for example).
@jhirono For the mult-agent example, I hope you have seen the second round of the MarLo competion on crowdai.org?
@ChristineTzeng Hi, We've had several reports of this happening but have not reproduced it. Could you do a help() on the env that you created to see what it contains (before you reset() it, say)?
something like this
INFO:marlo.base_env_builder:Agent missed 5 observation(s).
INFO:marlo.base_env_builder:Agent missed 1 observation(s).
INFO:marlo.base_env_builder:Agent missed 3 observation(s).
INFO:marlo.base_env_builder:Mission ended: out_of_time
INFO:marlo.base_env_builder:Mission ended: out_of_time
WARNING:marlo.base_env_builder:Error on attempting to start mission : There are not enough clients available in the ClientPool to start this 2 agent mission.
WARNING:marlo.base_env_builder:Will attempt again after 3 seconds.
WARNING:marlo.base_env_builder:Error on attempting to start mission : Failed to find an available client for this mission - tried all the clients in the supplied client pool.
WARNING:marlo.base_env_builder:Will attempt again after 3 seconds.but obviously, I should not run out of time, because in the training hours, common situation is that mission ends with command_quota_reached.
It seems like the minecraft is a little unstable, because I also trained with a total different code in a different computer and met the same problem.