Reproduce “Latent-Foresight: End-to-End Learning Predictable Representations for Latent World Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Efstathios Karypidis, Spyros Gidaris, Nikos Komodakis
Latent-Foresight predicts future road-scene features for downstream perception by jointly training a tokenizer and a flow-based predictor, rather than training the compressor separately. On high-resolution Cityscapes, its mid-term movable-object mIoU is 59.92, versus 58.05 for a two-stage autoencoder, with equal reconstruction cosine similarity of .991. Short-term all-class mIoU is effectively tied, and the evaluation does not test safety or actions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.