Associate Professor, MIT EECS
I lead the
See
Thoughts on research, teaching, and AI. View all posts
Computer vision conventionally maps images to intermediate representations (class, segmentation, 3D reconstruction, etc), but the "real" role of vision has always been as part of a perception-action loop. The LLM moment for vision won't mean SOTA on intermediate tasks, but instead intelligent, embodied agents. World models are the first glimpse. 3D in particular will become obsolete for training embodied intelligence models.
There are two disparate skillsets required for good research: moving fast to find what works, then slowing down to prove it. Knowing when to switch between them is crucial.
Courses I have taught at MIT.
Selected talks and presentations.
Vincent Sitzmann is an Associate Professor at MIT where he leads the Scene Representation Group. He received his PhD at Stanford University with Gordon Wetzstein and his Bachelor's degree from the Technical University of Munich. Vincent's research goal is to build machines that learn to understand and interact with the world autonomously by building ``world models'' - mental simulators that enable an agent to simulate what will happen next in their environment and as a consequence of their actions, a critical piece of AI and key in downstream applications in graphics, vision, and robotics.
