core-foundation

name	core-foundation
description	Always-on operational baseline for SpaceMind. Defines shared coordinates, action policy, safety guards, observability rules, and perception workflow for all tasks.
metadata	{"skill_kind":"core"}

Body Frame

• x: forward positive, backward negative.
• y: right positive, left negative.
• Use only the axes that are actually exposed by the current platform and tool profile.
• When dz, dpitch, or droll are not available or not meaningful on the current robot, do not invent vertical or tilt strategies.
• When geometry tools are exposed, interpret them in the same body frame as set_position(dx, dy, dz).
• The current image-side bearing of the target is not the same thing as the final task goal.

Attitude Frame

• dpitch < 0: look down (tilt nose down).
• dpitch > 0: look up (tilt nose up).
• dyaw > 0: turn right (rotate clockwise when viewed from above).
• dyaw < 0: turn left (rotate counter-clockwise when viewed from above).
• droll: generally not used for target tracking; prefer pitch and yaw.
• Attitude changes are relative deltas applied to the current orientation, not absolute angles.
• Typical per-step attitude adjustment: 10-30 degrees. Avoid jumps larger than 45 degrees in a single step.
• Use only the attitude axes that are actually supported by the current platform.

Motion Rules

• If target x > 0, the target is ahead, so approaching requires dx > 0.
• If target x < 0, the target is behind, so approaching requires dx < 0.
• If target y > 0, the target is to the right, so moving toward it requires dy > 0.
• If target y < 0, the target is to the left, so moving toward it requires dy < 0.
• Use z only when the current platform really supports meaningful vertical motion.
• Negative dx while target x > 0 is retreat, not approach.

Action Policy

• Call exactly one tool per step.
• Use set_position() for translation and set_attitude() for viewpoint control.
• Prefer a sense-then-act rhythm instead of repeating sensing or action without new evidence.
• If only raw image input is exposed, infer progress from target visibility, size, centering, cropping, and scene change rather than pretending you have metric truth.
• If the target is off-center or about to leave the frame, prefer set_attitude() before a larger translation.
• If the currently exposed tools do not include a capability, do not invent it in text; work with the available tool set.
• When the mission goal is satisfied or the mission should stop safely, call terminate_navigation(reason).
• Do not call the same sensing tool more than two consecutive times without an intervening motion or termination action. If a sensing result has not changed, act on it instead of re-reading it.

Optional Compute Tool

• If execute_code() is exposed, use it only when geometry, interpolation, threshold logic, or compact diagnostic reasoning genuinely benefits from computation.
• If you call execute_code() in this step, do not also call a motion tool in the same step.
• Let the code produce reusable output for later reasoning, not a trivial wrapper around an obvious answer.

Safety Rules

• When the target is already close, decide whether the mission is complete before taking another forward step.
• Do not terminate from a stale estimate. Require a fresh measurement or strong fresh visual evidence near the goal.
• If a fresh measurement shows x < 0, treat it as overshoot, not success.
• When uncertain, choose the safer action that preserves observability and avoids collision.
• Avoid repeating the same action type more than three times without evidence of progress.
• Do not continue forward blindly after the target is lost.

Progress Rules

• If recent steps are not reducing uncertainty or improving geometry, change strategy instead of repeating the same move.
• If repeated forward motion worsens observability, recover the target before continuing.
• For inspection-style tasks, avoid unnecessary translation when a viewpoint change or local visual tool can answer the question.

Observability Rules

• Keep the target observable while moving; do not let large blind motions accumulate.
• If the target is drifting out of frame or poorly centered, prefer set_attitude() before making a larger translation.
• If the target is temporarily lost, use the last known direction and the currently visible context to recover it conservatively.
• After one or two failed recovery attempts, re-sense or change recovery direction instead of repeating the same motion.
• If the target remains unavailable after repeated structured recovery attempts, terminate with a target-lost or observation-failed reason instead of wandering indefinitely.

Perception Workflow

• Use visual helper tools only when they are actually exposed in the current tool profile.
• Prefer local observation tools before moving when the needed evidence is already visible but too small or unclear.
• If no helper perception tools are exposed, rely on the raw current image plus recent action history and remain conservative.