Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
    SP Mohanty
    @spMohanty
    i just noticed that
    installed with v1.3 and it worked
    Best of luck :tada:
    parilo
    @parilo
    Thank you very much!
    Anssi
    @Miffyli
    Got an answer from Unity: Submission has to be Top50 and pass average 5 levels to progress to Round2
    Anssi
    @Miffyli
    Headsup for everyone: I am not 100% sure if this is required for Round1, but note that Rules section 6. asks you to submit your team name and team members to specific email.
    parilo
    @parilo

    Changed v1.2 to v1.3 and constantly getting

    ['/srv/conda/bin/python', '-m', 'pip', 'install', '-U', '-r', '/home/aicrowd/condaenv.q5kvyaz7.requirements.txt']
    Pip subprocess output:
    Collecting git+git://github.com/Unity-Technologies/obstacle-tower-env@v1.2 (from -r /home/aicrowd/condaenv.q5kvyaz7.requirements.txt (line 8))
      Cloning git://github.com/Unity-Technologies/obstacle-tower-env (to revision v1.2) to /tmp/pip-req-build-8xy5or5q
    Collecting scipy (from -r /home/aicrowd/condaenv.q5kvyaz7.requirements.txt (line 1))
    
    Pip subprocess error:
      Running command git clone -q git://github.com/Unity-Technologies/obstacle-tower-env /tmp/pip-req-build-8xy5or5q
      Running command git checkout -q dfe79a532cef4eaf1f320c451936c0b48a34d093
    ERROR: Could not install packages due to an EnvironmentError: [Errno 13] Permission denied: '/home/aicrowd/.cache/pip/wheels/b2/bb/cb/583bea9449bb9f042b43ec17678e9a2574a996f609fcbed317'
    Consider using the `--user` option or check the permissions.

    during 3 last hours, seems that aicrowd-repo2docker works not deterministic because after reverting back still getting this error

    Made one more submission but think that it will fail and not because of my fault
    and now it seems that there is a looong queue of submissions. If possible it would be nice to account submissions (let to resubmit) which was failed because of strange bugs in repo2docker or in some other part of the system
    parilo
    @parilo
    Ok, it failed again
    TashaSkyUp
    @TashaSkyUp
    Quick question... does our agent need to learn how to stand/walk/run/jump/turn etc?
    or is this more of a higher level challenge "figure out that keys go in doors" type of thing?
    KarolisRam
    @KarolisRam
    Higher level. the possible actions are jump, move left/right, move front/back, turn left/right and all combinations of them.
    Although I like the idea, the next level will be to have a humanoid robot learn to walk/jump from joint movements AND solve the obstacle tower!
    Arthur Juliani
    @awjuliani
    @KarolisRam that would be the ultimate challenge :P
    Anssi
    @Miffyli
    Or how about: A humanoid robot in front of virtual keyboard, mouse and monitor. The observation is image feed from the eyes of the humanoid, and task is to play OT by moving mouse/using keyboard with humanoid's hands? :derp:
    Maybe then we would be close enough to call it "fair playground" when evaluating against human players
    KarolisRam
    @KarolisRam
    Humans have the advantage of understanding how laws of physics work and not having to learn from scratch every time though.
    I'm working on getting Obstacle Tower into speedrun.com, hopefully it gets accepted soon and we can show what humans are capable of :) I doubt any of the Round 2 agents will get anywhere near human performance.
    Anssi
    @Miffyli
    Ah yes. I can already see somebody using the cube to trigger 20x level transitions and completing any% in five minutes :derp:
    KarolisRam
    @KarolisRam
    haha so true! Strafe jumps are possible to skip some gaps, not sure if strafe run is faster or not.
    SP Mohanty
    @spMohanty
    There are two common issues with the submisions over the last couple of days.
    Please refer to these two posts for solutions for the same :
    Anssi
    @Miffyli
    Anybody else noticing some weird "stuttering" of the character when you 'enjoy' your agent's play (i.e. let agent play and set realtime=True in env)?
    KarolisRam
    @KarolisRam
    yup
    Anssi
    @Miffyli
    Ah, should have checked the issues first :'D. Thanks!
    Anssi
    @Miffyli
    Pfft apparently seed 60, floor 11 is missing the target platform for the cube :D
    Anssi
    @Miffyli
    Looks like something is wonky on the evaluation server (getting The Unity environment took too long to respond all the sudden, and submissions seem to be failing for everyone)
    @spMohanty @awjuliani ^
    Anssi
    @Miffyli
    Hmm seems to be running for other people, but I am still getting that error message. I am using same submission code that worked a week ago.
    Arthur Juliani
    @awjuliani
    Very strange. Mohanty can say if this is due to something on the evaluation server side.
    I'll also look into seed 60 floor 11!
    Anssi
    @Miffyli
    Hmm seems like it now works correctly when I try it again on my Windows machine (seed=60 and skipping to floor 11) @awjuliani
    Could have been a momentary derp at some end :D
    Anssi
    @Miffyli
    @KarolisRam Silly note regarding speedrunning obstacletower: Looks like many agents learned to aim for the side of the exit door rather than going inside the exit room itself, thus saving split of a second :D . Reminds me of some speedrunning tricks with glitchy triggers and whatnot.
    KarolisRam
    @KarolisRam
    I noticed that! It would be fun to see if floor 15+ agents learn strafe jumps over gaps that are not supposed to be jumped over :)
    Or any other fun bugs
    Anssi
    @Miffyli
    Or using the cube like in some Half-Life speedruns to glitch through walls and whatnot :D. Maybe with enough brute forcing it could find those
    I wonder if using something like Go-Explore on "simpler" speedrun games (e.g. Mario 1) would eventually find those hard glitches and whatnot
    KarolisRam
    @KarolisRam
    I tried hard to find glitches in the environment myself, could only find the strafe jump. I've seen ML agents running super mario find bugs like jumping up from walls or killing goombas in weird ways.
    It's always fun seeing unintended solutions pop up in different ML problems. You can almost see some sort of milder paperclip maximizer scenario happening in the future :)
    Anssi
    @Miffyli
    Modern games tend to be boring glitch-wise with crashing most of the time, unlike the older console games with all kind of fun things :(
    Arthur Juliani
    @awjuliani
    There were some fun bugs that we found in Obstacle Tower earlier on, such as glitching through doors, climbing onto the "roof," etc. Glad to know that by the end of Round 2 there wasn't so much to exploit :D (sorry Anssi)
    Anssi
    @Miffyli
    Pfft yup, I noticed the clips/issues people posted :D. Can you enlighten us how those happened btw? No offence, but you would think spawning a (physics) cube inside a room wouldn't be that difficult with modern tools and especially Unity ^^
    Or the whole skipping-floors-every-now-and-then thing. I tried diffing the "de-assemblied" binaries, but it resulted into too much garbage I couldn't spot the fixing changes between versions
    Arthur Juliani
    @awjuliani
    Issues mainly arose from physics behaving differently at different time scales.
    Anssi
    @Miffyli
    Aaah gotcha. The curse of the modern games. Especially fun with locked framerates and whatnot :D
    Arjun Nemani
    @nemani
    Anssi
    @Miffyli
    Ah yes, the good ol' "Taught robot to reach goal in minimum number of steps. It decided to jump off the table" stuff :D
    hamid2026
    @hamid2026
    Hi, Is there any document or wiki to train the brain on the local computer?
    1 reply
    Shivam Khandelwal
    @skbly7
    Hi @hamid2026 & all,
    AIcrowdHQ has moved to Discord for IM connect.
    Join link: https://discord.gg/3jwn25E