Skip to main content
XR-1 RoboCasa365 predicts a 16-step action chunk from three camera histories, a state history, and a language instruction. It is the single-arm mobile-manipulation variant of XR-1. This checkpoint’s 60-wide output is padded for a common model representation. It does not expose 60 robot actuators. Use the action layout for this checkpoint; other robot embodiments require their own validated checkpoint and mapping. The first-actions example runs without a simulator, hardware, or local model weights. The simulation guide links the upstream benchmark and explains how the local-runtime adapter differs from a hosted bridge. A hosted simulator rollout still requires integration and validation. Reactor supplies inference; simulator execution and any physical robot controller remain client responsibilities.