Long-WAM scales world-action model context for real-time robot control
Long-WAM is a model-system framework for scaling the context of causal world-action models under real-time robot-control constraints; its authors report that longer histories only pay off when the video foundation is pretrained autoregressively. On RoboCasa GR-1, raising context from 0.0 to 19.2 seconds lifted success from 63.3% to 78.7%, while a bidirectionally pretrained initialization showed no net gain, and robot-domain autoregressive pretraining further raised peak success on GR-1 and LIBERO-Long. The paper also reports deployment on RTX 5090, DGX Spark and Jetson AGX Thor, with each action chunk including future-video latent prediction taking 107.4 ms on RTX 5090, plus real-time long-horizon manipulation on Unitree G1 and YAM reaching 95% success on dynamic cup stacking, where it says Pi0.5 and Fast-WAM succeeded in none of 20 trials.
Robot world models face a trade-off between long-horizon memory and real-time inference; this paper offers the testable claim that autoregressive pretraining is what makes long context useful, alongside an on-device deployment path.