This is the accompanying video for the paper “Open-Ended Learning Leads to Generally Capable Agents” (DeepMind 2021). All results shown are with the same agent and on hand-authored probe tasks that were held out of training.
Further reading blog:
Paper:
Timestamps of Results:
00:00 Intro
00:24 Tag Fiesta
01:23 King of the Hill
02:18 Hide and Seek
04:22 Capture the Flag
06:57 Catch ’em All
07:53 Choose Wisely
08:47 Nowhere to Hide
09:33 Stop Roll
10:43 Near not Near
11:53 See not See
12:42 Build Ramp
14:26 Outro
source
