Advances in AI at NYU - August 25, 2026

Join our virtual meetup to hear talks from experts on cutting-edge topics across AI, ML, and computer vision.

I'll present some of our ongoing work on using computer vision to create impact in the sciences. These target a two areas, solar physics and evolutionary biology, that deal with objects of radically different sizes but are unified by a need for high quality, trustworthy data.

I'll show off our efforts, done in collaboration with domain experts, that aim to produce the best possible maps of the Sun's powerful magnetic field and have created some of the world's largest repositories of data about bird morphology.

This talk will introduce Solaris: a multiplayer video world model in Minecraft. I will first present SolarisEngine, the software platform we built to simulate realistic multiplayer gameplay between bots at scale, enabling us to collect a large training dataset of aligned multiplayer actions and frames.

I will then discuss our staged training pipeline, starting with single-player pre-training before converting the model into a long-horizon multiplayer generator through bidirectional training, followed by causal training, and concluding with Self Forcing. I will also cover our memory-efficient implementation of Self Forcing, called Checkpointed Self Forcing.

Finally, I will showcase generated videos illustrating how Solaris maintains coherent long-horizon multiplayer interactions.

Collecting task-specific robot data for multi-fingered hands is challenging due to the many difficulties that arise in teleoperation. That is why recently there has been a major focus on learning robot policies directly from human demonstrations. However, human demonstrations are difficult to work with; there is a major morphological and visual gap between human and robot hands, as well as between the environments they operate in.

In this talk, I'd like to discuss my efforts on closing this gap.

Give an intelligent system a pile of broken pieces. Can it recover the object they once formed? And what if some pieces are missing? In this talk, I will present our recent work, from GARF to CRAG, on learning to reconstruct fragmented objects. GARF focuses on generalizing reassembly from synthetic training data to complex real-world fractures through large-scale fracture-aware pretraining and SE(3) flow matching. CRAG goes a step further by asking whether a model can imagine the whole in order to better assemble the parts. By jointly reasoning about fragment poses and complete object geometry, it uses global shape context to resolve ambiguous arrangements and recover missing geometry. Together, these works move 3D reassembly from matching visible pieces toward reasoning about the complete object behind incomplete observations.