Reproduce “Long-WAM: Scaling the Context of World-Action 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
Wei Huang, Bohan Zhang, Chenzhi Liu, Isabella Liu, Shuai Yang, Weian Mao, Luozhou Wang, Yicheng Xiao and 8 more
Long-WAM trains a robot controller to predict future video from visual history, then uses those predictions to generate actions. On RoboCasa GR-1, robot-domain autoregressive initialization raises success from 66.3% with 2.4 seconds of history to 78.7% with 19.2 seconds, while bidirectional initialization stays near 62%. This simulated comparison supports, but does not prove, the explanation that autoregressive pretraining helps use history.
Key insights
- Context length alone is not enough: robot-domain autoregressive pretraining yields larger gains from longer history than the tested bidirectional initialization.
- Useful context varies by task: RoboCasa GR-1 improves with history up to 19.2 seconds, while LIBERO-Long reaches 99.5% success with 2.4 seconds.
- Future video prediction can aid control without decoding predicted frames into pixels, though excessive context can hurt when it extends beyond the training data’s history coverage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.