Running Episodes

pyrualean play runs one episode: it boots the robot, gives the agent the task and serves its code until the episode is over.

pyrualean play --arm cells --suite libero_spatial_swap --task 7 --seed 2 \
    --model gpt-6-astra --max-turns 40 --budget-s 7200 --out runs/demo --cuda-device 0

--robot picks the robot (libero, the default, robotwin or robocasa) and --suite, --task, --seed the task instance; each robot page says what they mean there. Always pass --seed: play sends 0 when it is left out, whatever the robot's own default. RoboDojo runs through XPolicyLab (RoboDojo).

Cells or programs

--arm	What one LLM call runs	What carries over
cells	a cell in a persistent namespace with robo, like a notebook	variables, functions, imports, files
program	a complete program, def run(robo): ..., in a fresh namespace	files in the workspace, and the conversation

--max-programs N caps the number of programs on the program arm; 1 is a single shot whose result the agent still sees.

What the agent sees after a call

Option	Values
--images	on-motion (default): the current camera images after every call that moved the robot. on-demand: an image only when the code asked for one with robo.show(camera). none.
--feedback	rich (default): what the code printed, plus the robot's state and the primitives it called. pure: only what the code printed.

Budgets

Option	Meaning
--max-turns N	A decision budget: at most N code cells (python or run_program calls; finish is not counted). The agent knows the cells are limited, not how many.
--budget-s S	A wall-clock budget in seconds, thinking included (default 1200). With --max-turns it is a safety ceiling: set it well above what the cells take, so that outcomes do not depend on how fast the provider is (the paper's runs used 7200).

VLA policies, guides and shared servers

• --primitives no-vla removes the robot's VLA primitives: they disappear from the API reference and the knowledge text, and calling one is refused.
• --guides gives the agent RPent's three LIBERO operating guides in its workspace (optional reading); --guides-dir <dir> gives it the *.md files of another directory.
• --vla-endpoint host:port (and --sam3-endpoint on LIBERO-PRO) attach the episode to running model servers instead of booting its own; the simulator is always the episode's own.

A single shot

The strictest setting: one reply, one program, no feedback.

pyrualean card --suite libero_object_swap --task 2 --seed 0 --out runs/card --cuda-device 0
pyrualean generate --card runs/card/card.json --out runs/gen --model gpt-6-astra
pyrualean run --policy runs/gen/policy.py --suite libero_object_swap --task 2 --seed 0 \
    --out runs/code --cuda-device 0

card boots only the simulator and saves the task card; generate asks the model for policy.py (--transfer tells it the program will run on other seeds, so it must not rely on pixels); run runs the file once, with a hard timeout, and a crash or timeout counts as a failed episode.