Reproduce “AR-WAM: efficient streaming world-action modeling for autonomous driving” 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

Baiting Luo, Shatadal Mishra, Ahmadreza Moradipari, Nejib Ammar

AR-WAM combines a reusable cache of causal visual history with a single-pass prediction of noisy future frames, whose features a trajectory expert uses alongside cached observations. On NAVHARD, it scores 38.6, versus 37.3 without future conditioning and 38.8 with ten passes, at 1.0× rather than 4.2× latency. This result is limited to the reported variants and evaluation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.