Start with a single Synthetic episode, then install one robot environment:
Each page describes the isolated interpreter, pinned sources/assets, discovery checks and a one-task run. Run one task end to end before you run a suite. RoboDojo's default policy is code mode, so a RoboDojo host also needs bubblewrap, unprivileged user namespaces and the sandbox interpreter (code-mode sandbox); the LIBERO-Pro and RoboTwin defaults use tool calls and need none of them.
examples/official_suite expands the LIBERO-Pro and RoboTwin suites from the current default composition into cells, one episode on one seed or initial state each. Official seed selection, task/source versions, episode budget and environment verdict stay explicit in every cell. ./scripts/ora suite --help runs composed cells on the simulator slots of your host profile that you name, one cell per slot at a time, and does not launch on a slot that another run holds.
An agent that acts through robot tools is not comparable one-to-one with a trained VLA policy. When you report results, state its observation and action interface, the task and seed protocol, the attempts and the budget. A passing installation or qualification check shows that the setup works; it does not measure the model or establish benchmark success.