worldmodels.fyi
← Catalog
PaperTsinghua University · 2024

iVideoGPT: Interactive VideoGPTs are Scalable World Models

A scalable autoregressive transformer world model with a compressive tokenizer, pretrained on large video corpora for interactive prediction and control.

Summary

iVideoGPT casts world modeling as interactive video prediction: a compressive tokenizer plus an autoregressive transformer predict future frames conditioned on actions, and the model scales via pretraining on millions of trajectories. Fine-tunes to action-conditioned prediction, model-based RL, and visual planning.

Metadata

Authors
Wu et al., Tsinghua
Year
2024
arXiv
2405.15223
Venue
NeurIPS 2024

Relationships

Introduces

iVideoGPT
#autoregressive#video#model-based-rl