Open X-Embodiment / RT-X

Adopt with guardrails

Why

OXE/RT-X is proven foundational infrastructure — it is the actual training corpus behind OpenVLA (970k demonstrations) and the direct methodology ancestor of modern generalist manipulation policies like Physical Intelligence's pi-series — but the collaboration's own paper documents real data heterogeneity and model-capacity limits, and the maintaining GitHub repo now sees near-zero commit velocity, so treat it as a mature reference dataset to build on rather than an actively evolving platform.

Next

Pull the relevant Open X-Embodiment subsets via the RLDS/LeRobot loaders, budget engineering time for camera-pose and action-space renormalization to the target robot's own hardware, and validate transfer empirically per-embodiment rather than assuming it — the collaboration paper itself provides no predictive criterion for when positive transfer occurs.