Nvidia Researchers Release NitroGen Trained To Play 1000+ Games
A team of researchers from Nvidia, Stanford, Caltech and collaborators this week introduced NitroGen, an open-source foundation model trained to play 1,000+ video games using over 40,000 hours of public gameplay video and streamer-input action overlays. Built on the GROOT N1.5 architecture, NitroGen shows strong cross-domain competence—achieving a 52% relative improvement in task success on unseen or procedurally generated games versus models trained from scratch—and the code, weights, and dataset are openly released.
Key Points
- 1Trains on 40,000+ hours of public gameplay to play more than 1,000 diverse video games.
- 2Demonstrates 52% relative improvement on unseen procedural games versus scratch-trained models.
- 3Enables open-source reuse for simulation, robotics, and embodied agent research and fine-tuning.
Scoring Rationale
Significant open-source embodied-model advance with clear robotics potential; limitation: early-stage results focused mainly on fast motor control.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
