NVIDIA Cosmos powers 4 of top 5 winners in ECCV AI City Challenge forecasting track

NVIDIAAI · x · 2026-09-09

NVIDIA recaps the AI City Challenge at ECCV 2026, where teams built visual AI systems that go beyond frame-level recognition across six tracks and two out-of-domain leaderboards — handling multi-view scene understanding, safety event explanation, visual evidence retrieval, prediction, and adaptation.

Notably, four of the top five solutions in the Generative Video Forecasting track (Track 5) used variants of NVIDIA Cosmos world foundation models. Public-leaderboard winner Qyn fine-tuned Cosmos3-Nano into CosmosAlign to forecast realistic traffic scenarios from short visual histories and motion descriptions.

Original post →

More from Embodied

Embodied channel →