Thursday, 8 October 2026 · 6 min read
Long-WAM improves robot control by scaling visual history with autoregressive video pretraining
Long-WAM is a robot control framework that leverages long visual histories effectively by first autoregressively pretraining video prediction models, then adapting them for action generation, achieving up to 78.7% success on RoboCasa GR-1 with 19.2 seconds of context and real-time deployment on various hardware.
Based on the paper Long-WAM: Scaling the Context of World-Action Models by Wei Huang, Bohan Zhang, Chenzhi Liu, Isabella Liu and 12 others
Challenges of using long visual history for real-time robot control
Effective robot control requires understanding not just the current scene but also how it is evolving over time, which demands access to visual history. However, processing longer histories can introduce delays that reduce responsiveness, creating a trade-off between richer context and timely action.
World-action models (WAMs) integrate video prediction with action generation to leverage visual history for control. Prior work retains history via caches or memories but often does not distinguish between merely accessing history and effectively using it for prediction.
The key question is how increasing the length of visual context affects control performance under real-time constraints, and how pretraining methods influence the ability to exploit longer histories.
Sign in to read the whole paper
You have read the opening. The rest is free with a Bezen account, and no card is needed.
In the rest of this paper
- How the method works, step by step, with a flow chart
- Three key formulas from the paper, explained symbol by symbol
- Three figures from the paper, and how to read them
- How it was tested, what was found, and where it falls short