AnimalAI - v3.1.2.exp

[alhasacademy96]

Introducing AAI v3.1.2.exp[erimental].

This build is in experimental status so not everything might work or corrected (regarding the changes detailed below:).

• Fixed the Spawner Tree "clock" discrepancy,
• Introducing the final version of End-of-Episode Notification feature, with a brand new look (with sound effects added).
• Modified Spawner Pillar Button to make it more robust to handling missing rewards, weights and probability values.
• [New Feature] Camera and Reset buttons now toggled on/off by specifying in the yaml file.
• Cleaned relevant scripts to the above additions/modifications for improved code readability and functionality.

Steps:

-Download AAI-3 OS build for your system from below
-Extract files anywhere you'd like on your system (we use Winrar to package application contents)
-Run .exe/.app file (depending on your system) in the extracted folder of AAI-3
-Enjoy!

---

N.B: If you are using Linux, you may need to make the .exe file executable: Simply run this command in your terminal: chmod +x env/AnimalAI.x86_64. If you are using MacOS, you may also need to run this command: chmod -R 777 [applicationName].app in your MacOS terminal to unlock permissions for running the application. Edit: Please also make sure when you extract the folder, you move the files inside the sub-directory to its parent folder.

What's Changed

• Adapt environment to allow changes to timescale and framerates for more control by @kozzy97 in Adapt environment to allow changes to timescale and framerates for more control #28
• Improved raycastparser.py by @alhasacademy96 in Improved raycastparser.py #29

Full Changelog: v3.1.0...3.1.2

---

This discussion was created from the release AnimalAI - v3.1.2.exp.

[kozzy97]

Hi @alhasacademy96, thanks for pushing this update. I'll write any issues I find with this build here.

[kozzy97]

I've been playing with the spawner tree at different timescales and everything works perfectly fine at different timescales, which is great!

[kozzy97]

Although things do break above timescale=100. I tried the old training default of timescale=300, and I can't see anything spawning, but that might be because it's not rendering well in the full display at such quick timescales. Not sure how to check that.

[kozzy97]

I've been playing with the end-of-episode notification using this configuration:

!ArenaConfig
arenas:
  0: !Arena
    pass_mark: 0
    t: 250
    canResetEpisode: true # allow the user to change the perspective. Default is true.
    canChangePerspective: true # allow the user to change the perspective. Default is true.
    showNotification: false # show/hide the notification box. Default is false.
    items:
    - !Item
      name: Agent
      positions:
      - !Vector3 {x: 20, y: 0, z: 20}
      rotations: [0]
    - !Item
      name: GoodGoal
    - !Item
      name: BadGoal
    - !Item
      name: GoodGoalMulti

Here are some thoughts:

1. The notification and sound work very well, although we should talk with the rest of the team about exactly what image and sound to use.
2. When I get the notification, the instance continues playing. For example, I can go and hit the BadGoal, get the notification, but then I can go and get the GoodGoalMulti and increase my score. We need the episode to end as soon as you obtain an episode-ending goal (or you collect all the available GoodGoalMulti). That works fine when the notification is switched off.
3. I still think that somehow having the toggle in the python API is superior to having it in the yaml, because if you want notifications, you probably want them for all the configs defined in a single file. People who have built configs already without this feature might have to go back and add extra lines to thousands of configs, which isn't ideal. But I realise that this might be difficult, so the yaml way of doing it is a good interim solution, but ultimately it would be good to be able to do it both ways.

[alhasacademy96]

The gifs and sounds are really a placeholder. i just wanted to demo that its possible to play any video and sound in the notification itself :)

[kozzy97]

I've just played with canChangePerspective and canResetEpisode. Here are my thoughts:

1. They work perfectly well and as described.
2. I think canChangePerspective default perspective should be first-person in the final release, as that is what the agents get in testing. It doesn't really make sense otherwise.
3. Users should be able to specify which perspective they want to have as default. Currently, when canChangePerspective = true then it is birds-eye view, and when canChangePerspect = false it is third-person view. Keen to hear if this is possible for the next build.
4. As with the notification, I would strongly prefer this to be definable python-side rather than in the yaml, or ideally to do both. But I realise that this might be a future version update.

[kozzy97]

I've just played with Pillar-Button. Here are my thoughts:

1. Most of the parameters seem to work well.
2. It is detected by raycasting.
3. Randomising the button position works perfectly.
4. Randomising the goal spawning position does not work. That's fine, but should be flagged. I tried with this setting rewardSpawnPos: !Vector3 {x: -1, y: 0, z: 24} and this setting rewardSpawnPos: !Vector3 {x: -1, y: -1, z: -1}.
5. @wschella 's issue here is still present, although slightly different. When rotation: [0] you get the expected behaviour. When rotations: [90] you get the button laying flat on the ground (not in the ground like this). You can still press the button even when it's on its side. When rotations: [180] the button disappears completely. When rotations: [270] the button is on its front on the ground, and you cannot press it. If possible, rotations should change its rotation on the y-axis.
6. I also find it a bit strange that the button's initial rotation is always facing the west wall (from birds eye view). I don't know why, but my default is that it should face south wall, but that might just be my bias.
7. I have played with spawning it on top of blocks. It works and drops from a height. An interesting behaviour to note is that if a wall is 5-units high and you try to spawn the pillar button at y: 5, it won't spawn. It only starts spawning at y: 5.8 or above, which seems kind of arbitrary. It still drops a little bit from a height there. Any idea why there's that behaviour?

Overview of issues, in order of priority (imo):

• Rotation behaviour is weird.
• Strange height limit when spawning on a block.
• Randomising goal spawning position (if possible, but not super necessary).

[kozzy97]

That concludes my testing, along with the issues logged here and here. Let me know if anything is unclear @alhasacademy96 . Thanks!

[alhasacademy96]

Awesome work on the feedback @kozzy97, i think we should defo keep this format of testing/feedback at least on github as we did for this experimental release. Some of the issues are already known but some such as the spawn of the rewards randomly should have been already working, where in this build it seems to be broken, which is fixable. :)

[wschella]

@alhasacademy96 Can it be there is no linux build attached this time? To test if the spawner tree issues are resolved, I'd need that.

[alhasacademy96]

Soooo sorry i forgot to upload the builds for the other OS'. Uploading by the end of today.

[alhasacademy96]

LINUX build now live

@wschella @kozzy97

[wschella]

The spawner tree behaviour seems fixed now, thanks!