Thank you @davidchalmers42 for featuring https://
minedojo.org in your talk. We build open-ended agents that can solve any task in Minecraft. This requires the agent to have awareness of the Minecraft world. Foundation models provide the agent with priors (e.g. reward function)
AGENTS
-
MineDojo: Open-ended Agents for Minecraft Task Solving
By
–
-
CASCADE: Learning World Models with Multi-Agent Reward-Free Data
By
–
Learning General World Models in a Handful of Reward-Free Deployments.
CASCADE seeks to learn a world model by collecting data with a population of agents. @YingchenX
, @JParkerHolder
, @AldoPacchiano
, @PhilipJohnBall
, @_Oleh
, Stephen J. Roberts, @_Rockt
, @EGrefen -
SAMPLR: Minimax Regret Method for Unsupervised Environment Design
By
–
Grounding Aleatoric Uncertainty for Unsupervised Environment Design. @MinqiJiang
, @MichaelD1729
, @JParkerHolder
, @_AndreiLupu
, Heinrich Küttler, @EGrefen
, @_Rockt
, @J_Foerst propose SAMPLR, a minimax regret UED method that optimizes the ground-truth utility function. -
Language Dynamics Distillation Improves Policy Learning
By
–
Improving Policy Learning via Language Dynamics Distillation. @hllo_wrld
, @Jayelmnop
, @LukeZettlemoyer
, @EGrefen
, @_rockt propose Language Dynamics Distillation (LDD), which pretrains a model to predict environment dynamics given demonstrations with language descriptions -
Language Abstractions Improve Intrinsic Exploration in AI
By
–
Improving Intrinsic Exploration with Language Abstractions. @Jayelmnop
, @hllo_wrld
, @RobertaRail
, @MinqiJiang
, Noah Goodman, @_rockt
, @EGrefen explore natural language as a general medium for highlighting relevant abstractions in an environment. -
Exploration Integral to General Intelligence and Learning Systems
By
–
General Intelligence Requires Rethinking Exploration. @MinqiJiang
,
@_rockt
, @EGrefen argue that exploration is integral to all learning systems despite the fact that the study of exploration in AI has mostly focused on reinforcement learning. -
AI Character Creation Platform Inspired by Character.ai Before ChatGPT
By
–
note: inspired by similar to something like http://
character.ai, but also came up with this a few weeks before chatgpt was released so a little foreshadowing there -
Update: ChatGPT external browsing works, but likes post as Grimezsz
By
–
Update — I got external browsing working and ordered ChatGPT to like this post, but for some reason it was logged into Twitter as @Grimezsz
: -
Building and Refining Your World Model Over Time
By
–
Yes, you definitely should have a world model, but you learn/refine it as you go along.
-
Improving Sequential Decision Making by Incorporating Backpropagation Principles
By
–
But this is well below that. Let me put it another way: how can we improve RL by making it more like backprop? (And by RL I mean sequential decision making, not the current set of techniques for doing it.)