A new technique makes robot control AI far more reliable when the world doesn't look like its training data.
Researchers built CSWAM, an upgrade to an efficient robot action model called FastWAM. FastWAM predicts robot actions quickly, but it stumbles when scenes look different from training footage, since its representations lean on visual appearance rather than what's actually happening in a scene. CSWAM adds a module built on a model called V-JEPA 2.1 that tracks state changes and motion over time instead of surface appearance, drawing on a short window of recent and current observations. That context feeds into both the video and action-planning parts of the system through causal attention, while the model still runs action-only inference at deployment, so it stays fast.
The gains are large: in simulation, success on a clean-to-randomized generalization test jumped from about 10% to over 45%, and across real-robot tasks with varying difficulty, average success rose from 27.5% to 70%. For an industry racing to put general-purpose robot brains into warehouses and homes, the ability to handle scenes that don't match training data is the actual bottleneck, not raw dexterity in curated demos.
Benchmark jumps like this are common in robotics papers and don't always survive contact with real-world clutter, but a 42-point swing across real robots, not just simulation, is harder to wave away as overfitting.