Breakthrough Minecraft General AI agent does over 3000 tasks the artificial intelligence research team behind the work states that GPT3 from open ai is powerful but blind they believe the future of foundation models will be embodied artificial intelligence agents that proactively take actions endlessly explore the world and continuously self-improve In their NeurIPS paper Mine-Dojo building open-ended embodied agents with internet scale knowledge they provide a blueprint for this future and what it will take the artificial intelligence researchers believe three main ingredients are required for generalist AI agents to emerge the first ingredient is an open-ended environment that allows an unlimited Variety of tasks and goals to be carried out an example of this would be planet Earth as it is Rich enough to forge an ever-expanding chain of life forms with different behaviors the second ingredient is a large-scale knowledge base that teaches an artificial intelligence agent Not just how to do things but also what are useful things to do GPT3 from open AI learns from web text alone but the team behind the Mind-Dojo a agent wanted to use much richer data like video walkthroughs multimedia tutorials and freeform Wikipedia information The third ingredient is an agent architecture flexible enough to try any task in open-ended environments and scalable enough to convert large-scale multimodal knowledge sources into actionable insights as something like an embodied GPT3 these ideas landed the team in Minecraft Saying the infinite voxel world is in a league of its own and never ceases to amaze unlike other games Minecraft defines no particular score to maximize and no fixed storyline to follow making it well suited as a truly open-ended AI playground the research team introduces Mine-Dojo which is a New open framework to help the community develop generally capable artificial intelligence agents Mine-Dojo features a simulator Suite based on Minecraft a massive internet database including YouTube Wiki and Reddit and a simple but promising Foundation model recipe for AI agents Mine-Dojo simulator unlocks the full potential of Minecraft for AI research and supports versatile Observations like RGB voxel radar and GPS as well as flexible actions like movement crafting inventory Ops and more terrains weathers and lighting are all highly customizable mine dodo has over 3000 tasks making it one of the largest agent benchmarks ever created The tasks range from being clearly defined like “sheer sheep to obtain wool” to more open-ended creative tasks like “play fireball with a ghast” and “build a two-story house with a swimming pool” Minecraft has an enormous online presence where over 140 million players produce a treasure trove Of knowledge every day crystallized into video streams Wiki pages and forum threads the researchers scrape from this collective wisdom and creativity to make the data accessible to everyone the first part of their database is YouTube where they collected over 300,000 hours of narrated gameplay videos with over 2 billion transcribed words timeline transcripts Enabled in artificial intelligence to ground free-form natural language in video pixels and learn about diverse activities without the need for human labeling the second part is Minecraft Wiki which covers almost every aspect of game mechanics and provides unstructured knowledge in multimodal tables recipes illustrations and step-by-step tutorials The researchers aggregated about 7,000 pages that interleave text images tables and diagrams the third is Reddit where they gathered 340,000 posts and 6.6 million comments in R/Minecraft people ask questions showcase cool builds and discuss General tips and tricks large language Models can be fine-tuned on their Reddit corpus to internalize Minecraft specific concepts and acquire new strategies finally the researchers propose a conceptually simple method to learn a Minecraft playing agent from in the wild YouTube videos it is far from solving the game but shows a Baby step towards the team’s vision of an embodied GPT3 that takes the right actions given any language prompts since the YouTube data set has time aligned narration they were able to train a video language contrastive model called Mine-CLIP which intuitively learns to associate a video With a text that describes the video activity and then computes a correlation score between 0 and 1. to use Mine-CLIP for training the agent takes a text prompt and interacts with the Minecraft synth to generate a video which can be fed to Mine-CLIP to compute its correlation with a prompt the Height of the correlation the more the agent’s behavior is on the right track they essentially repurpose Mine-CLIP to be a foundation reward AI model which provides a dense reward signal for any task described in open vocabulary English and can be plugged into any reinforcement learning Artificial intelligence algorithm they have chosen to open source everything including the simulation suite database algorithm code pre-trained models and even annotation tools next the artificial intelligence research team hopes it will continue to improve with the help of the community DeepMind builds interactive video game artificial intelligence Google DeepMind AI presents a Framework for developing artificial intelligence agents that can follow instructions from humans and operate in unstructured environments Google DeepMind AI has demonstrated the first steps in creating video game ai’s that can comprehend fuzzy human Concepts to begin interacting with people on Their own terms although it is still in its early stages this Paradigm develops agents that can talk listen ask questions navigate search for and retrieve information manipulate things and carry out numerous other tasks in real time in the Video Game World interactions between players are where The framework starts DeepMind provided agents with a wide-ranging yet unpolished range of behaviors through imitation learning it is essential to have this Behavior prior in order to allow humans to judge the interactions better agents are created with further to an evaluation of The agent’s Behavior and the optimization of these evaluations happens through reinforcement learning instead of increasing a game score DeepMind a had people create assignments for the AI to judge its progress and Improvement through this method this Paradigm from DeepMind enhances aged Behavior through open-ended and grounded human contact DeepMind used three steps to construct the AI agents they begin by teaching the agents to mimic the fundamentals of straightforward human interactions such as when one person asks another person to perform something or provides an answer This stage according to DeepMind entails developing a behavioral prior that enables agents to frequently engage in meaningful interactions with people without this imitation phase agents just move and speak arbitrarily which makes it nearly impossible to meaningfully connect the researchers use the reward model to optimize agents after training it on human preferences The agents were put into the simulator and given instructions to respond to questions and follow rules the trained reward model assessed the agent’s actions and speech as they occurred in the environment then utilized our reinforcement learning algorithm to maximize their performance First they reuse the tasks and questions from the human data set to get even more task guidelines and questions then two agents engage with one another with the blue agent mimicking humans by setting tasks and offering questions and the yellow agent obey instructions And responding to the questions DeepMind is hopeful that one day using the architecture gaining artificial intelligence will be able to naturally respond to humans rather than relying solely on pre-written behavioral patterns they noted that building everyday digital and robotic helpers for human interaction with might also be possible using the architecture Video Information
This video, titled ‘Nvidia Minecraft General AI Does 3000+ Tasks | NEW DeepMind Video Game Artificial Intelligence’, was uploaded by AI News on 2022-11-26 15:21:07. It has garnered views and [vid_likes] likes. The duration of the video is or seconds.
AI Marketplace: https://taimine.com/ Deep Learning AI Specialization: https://imp.i384100.net/GET-STARTED Breakthrough …