[TOC]

IGLU Minecraft challenge NeurIPS 2021 workshop

https://nips.cc/virtual/2021/competition/22444

The organiser introduced a Web environment for data collection

Architect task

Baseline model

image-20211210070437284

Ha! No winner that outperforms the baseline?!

image-20211210070601738

…. which means I have a chance !

Builder task

image-20211210070730935

image-20211210070753279

image-20211210071041719

image-20211210071153129

Their Future direction

image-20211210071234132

Breakout session

image-20211210072219363

image-20211210073221055

image-20211210073323939

image-20211210073440925

image-20211210073515210

image-20211210073601431

Winner Solution

image-20211210073944203

image-20211210074011100

image-20211210074034922

image-20211210074101709

image-20211210074138037

image-20211210074226923

image-20211210074253093

image-20211210074330927

image-20211210074349681

image-20211210074512479

image-20211210074553334

image-20211210074649058

image-20211210074719218

image-20211210074737579

https://github.com/LinarAbdrazakov/IGLU-Minecraft

Highlight:

the data augmentation method

they tried to change the colour of objects / translate position in the training dataset so as to increase the number of training data

Evaluation metric brainstorm discussion

https://docs.google.com/document/d/1e9zBuX50JDpuF9JBfKMRT1S9yXb_W4KLevH3PtmfYvM/edit

They are struggling to get a automatic evaluation metric that can evaluate actions that reflect high level instructions

image-20211210082931320

They freeze the Architect task since it seems too difficult to do.

They want to get more people to participate next year.