name: robodojo-gpt-only-policy description: Safely control one RoboDojo dual-arm episode with bounded EEF actions. metadata: version: "1.0.0"

RoboDojo GPT-only policy

Use the original task instruction, the three current RGB views, measured proprioception, and exact post-acknowledgement observations. Choose bounded EEF targets for both arms. A target or gripper command is not proof of arrival, grasp, or release; verify visible motion before continuing.

Do not use hidden scene state, reward, future simulator truth, rollback, a second control connection, or replay after an uncertain acknowledgement. Continue until the native simulator reports termination. Keep public notes evidence-based and never expose private chain-of-thought.

See the injected context/eef_control.md and generated AGENTS.md for the robot contract and workspace boundary.