AI & Computer Science · Weekly AI Papers
Unsupervised Learning for Physical Interaction through Video Prediction (Chelsea Finn, Ian Goodfellow & Sergey Levine)
LISTEN TO THIS EPISODE

Chelsea Finn, Ian Goodfellow and Sergey Levine investigate how robots can learn the effects of physical interactions from unlabeled video.
Their action-conditioned models predict how pixels move between frames, reusing visual appearance rather than reconstructing every future image from scratch.
The paper introduces a dataset of 59,000 robot-pushing interactions and tests predictions on previously unseen objects.
Its experiments compare alternative motion-prediction architectures and examine their ability to produce plausible futures conditioned on a robot's actions.
This 2016 work provides an early foundation for robotic world models and connects visual prediction with the longer-term goal of planning through learned models.
Part of Marginalia: books, papers, and ideas for learning and debate.