Topic 01 of 17
World models in robotics
A world model is a learned simulator: a neural network trained to predict what a robot will observe next, given what it observes now and what it does. Instead of learning a task by acting on real hardware, a policy can plan or even train entirely inside the model's predictions, which is dramatically cheaper and safer than real-world trial and error. The idea runs from Ha and Schmidhuber's 2018 "World Models" through the Dreamer line to today's video-scale models such as DreamerV3 and the generalist robot world models trained on human video.
arXiv query: cat:cs.RO AND (abs:"world model" OR abs:"latent dynamics" OR abs:"video prediction")
Presented at reading clubs
These already have a slide kit, so they are the fastest ones to pick up and present.
Recent work
294 papers in this corpus carry the world models tag. How tagging works.