manfredch/lyapunov/action-execution

Execute an action with an existing robot

Reuse the current scene and robot when the user requests an action. Stop directly with robot_stop/sim_stop; use policy_stop for policy Jobs and job_kill for long Jobs, then read their real terminal states. Do not wait for skill loading or planning before stopping.

1. Identify the instance from the current selection and scene_inspect. Use sim_world_list to find a real world in the same session and Scene. Call sim_open only if none exists, and sim_sync only if its version is behind. Retain the current worldId, generation, and appliedSceneRevision; do not guess between multiple candidates. A successful Scene edit does not establish that physics is synchronized.
2. Read real joints, actuators, units, TCP, base, and availability with robot_describe. Resolve only the current missing information. When a tool is unavailable or rejects the action, report the specific reason and next step.
3. Use the interface for this action, with the Scene revision/world generation and actionId required by its schema. Read the returned receipt, sim_action_receipt, or robot_state after execution. If the result is unknown, query the same action/Job before resubmitting motion that may already have executed.

Manual arm control does not require a trained policy. If TCP information is missing, choose a real end effector from robot_describe.nativeBodies/nativeSites, then save either a site or a body-local offset with robot_set_tcp. A standard Panda has no site; do not assume an official TCP. Use robot_set_base modes source/free/fixed and real world or entity-body anchors; parentId does not establish a weld. Small-step IK currently requires a fixed base. If free/entity-bound bases are unsupported, use the actually controlled joints for manual adjustment. The human entry is the robot manual controls panel, using the same Scene CAS, synchronization, and Frame path. Read real presets and missing information with robot_presets when it is visible.

Use motion_plan / motion_path when obstacle avoidance or a full trajectory is needed. Joint, TCP, and grasp descriptions do not extend policy compatibility. Match G1 12DOF and 23DOF/75-dimensional contracts only against the real robot and observation mapping. sim_reset stops and rebuilds all instances in the current world without changing the Scene.

Perform bounded readback for the current goal. Judge visual changes through a real observation from the same window and Scene version. Judge distance, standing, collision, support, and drift through engine readbacks from the same generation. Submitted actions, prepared files, and historical test results do not prove that this goal was reached. Workspace scripts follow existing permissions; product state is modified through its corresponding interfaces.

Do not edit product source, state databases, or resource directories directly to bypass an interface.