[Announcement] MineRL 0.2.0 + Baselines! UPDATE FOR FIXED DATA πŸ’―

#1

MineRL 0.2.0 released!

HEY GAMERS! Excited to announce a HUGE update for minerl and the dataset UPGRADE FOR FIXED DATA!

cropped_viewer

The package is now on minerl-0.2.0, and we have released a new version of the dataset! To upgrade:

pip3 install --upgrade minerl

Then get the new dataset by running the following in python3!

import minerl
minerl.data.download('your/dataset/path`) 

New in minerl-0.2.0:

# Make sure your MINERL_DATA_ROOT is set!
export MINERL_DATA_ROOT='/your/local/path'

# Visualizes a random trajectory 
python3 -m minerl.viewer MineRLObtainDiamondDense-v0
  • Fixes an issue where the camera values are outside the interval [-180, 180] (#66)
  • Fixes an issue where there were invalid enum fields in the craft actions (#67)
  • Fixes an issue where the compass angle of a demonstration in the navigate task was always non-positive. (#75)
  • Fixes an issue where NavigateExtremeDense-v0 had extreme rewads (#76)
  • Adds a suppressor for the BrokenPipe error in seq_iter (#83)
  • Deprecates seq_iter for a new function sarsd_iter which returns obs, act, reward, next_obs, is_next_done
  • Fixes an issue where trajectories would complete the challenge but additional frames would be in the data (#127)
  • Removes a corrupted data stream (tempting_capers_shapeshifter-14_…) (#130)
  • Fixes an issue with the data downloader where the read operation would time out (#131)
  • Adds resuming of downloads to the data downloader!
  • Fixes massive issue with data alignment!
  • Frames at the beginning of experiments with teleportation are (mostly) removed!
  • Fixes an issue where random rewards would appear during experiments (opening a furnace).

SUBMISSIONS SHOULD BE OPEN NEXT WEEK SO GET EXCITED! :100::100::100:

ChaineRL Baselines for Auxiliary Environments released!

The awesome folks over at PFN have released some initial baselines in their framework ChainerRL on some of the auxiliary envrionments (Treechop + Navigate)!

Check out their amazing work here!

They’ve released DDDQN, Rainbow, & PPO, and they solved MineRLTreechop! Feel free to ask them if you have any questions!

Note: when AIcrowd releases the competition submission and evaluation kit these baselines will be converted into that format an appear on the leaderboard. So check back for updates.

5 Likes
#2

Hey! This looks awesome!! :heart_eyes:

2 Likes
#3

We’re glad you think so! Excited for everyone to give it a go =)

#4

Hi! Whats the size of new dataset? Capture

It went till 95GB when I left it to download last night. I thought something went wrong so started the download again, but it exceeded the 15.10GB mark again. What’s going on here?

Is it possible to give a direct download link?

#5

Hmm this looks like an issue with the continuation after a failed download code - we will fix this asap. Meanwhile direct download is possible at https://router.sneakywines.me/minerl_v1/data_texture_0_low_res.tar.gz

3 Likes
#6

If you delete your cache folder (/temp/pysmartdl) does this fix the issue?

1 Like
#7

Yeah It fixed it. Thanks!

1 Like
#8

@BrandonHoughton Any progress on this?

#9

Now the chainerrl baselines do not use the dataset.
Will there be a version that use the dataset ? such as behaviour clone

1 Like
#10

Just to update this thread - the downloading issue has been fixed! Chainer baselines will be released for the data soon - the Preferred Networks team is working hard to ensure the provided baselines are high quality