version: "1.0.0" name: go2-pixelgoal description: Navigate a Go2 with four-way PixelGoal observations and bounded visual hops.

Go2 PixelGoal Operating Notes

• The harness attaches the initial front/right/back/left views in that order.

Inspect them, call exactly one pixelgoal_navigate(point=[x, y], direction=...) hop per observation, then repeat from the returned views until the goal is visually satisfied.

• Every navigation point must sit on a clearly visible, unambiguous surface.

Preferred: clearly traversable floor. Also allowed as a long progress hop: a large stable vertical surface in the travel direction (wall face, cabinet face, solid door panel, large furniture front) — the robot stops about 0.35 m in front of the marked surface, so a distant stable surface is a safe way to cross open space efficiently.

• When the travel direction is clear, choose the farthest clearly visible safe

option: the farthest traversable floor, or a stable vertical surface farther along the route when floor runs out of view. Use a shorter local hop only when farther choices are occluded, ambiguous, or unsafe.

• Never click glass, windows, mirrors/reflections, transparent doors or

anything seen through them; thin structures (chair/table legs, cables, bags, stands); the floor-wall boundary line itself; sky, ceiling, the horizon, or empty space. Keep a safety margin from obstacles and floor boundaries.

• For a find/search task, never click the target object itself; stop only when

it is close enough for human-verifiable identity, not merely visible far away.

• Keep the robot still between observation and navigation so the normalized

point still matches the latest camera frame used by PixelGoal.

• pixelgoal_navigate waits for the hop to finish and returns its terminal

status, camera freshness metadata, and fresh front/right/back/left views. Do not navigate from a missing or stale view. Call observe_surround or pixelgoal_status only to recover from missing/stale images or diagnose an abnormal result.

• The map backend can perform successive exploration legs of up to 1.5 m when

the target is unresolved, rechecking after each successful leg. The default limits are 60 seconds for the whole hop and a 10 m exploration budget. Exploration stays close to the clicked direction. At an incomplete map edge, it may extend at most 0.5 m into unknown space from reachable free space; another unknown leg requires new free-space evidence near that frontier. This does not permit clicking visually unsafe or ambiguous surfaces.

• needs_agent / exploration_complete is an intentional observation handoff

when the exploration budget or reachable free space runs out, not a failed tool call, a request to keep polling, or proof that the point/task was reached. Inspect all fresh returned views and choose a new point if movement is needed.

• timeout / deadline_exceeded ends the entire hop; inspect its fresh views and

choose the next safe point if more exploration is needed. Do not resume the old hop or automatically retry its point without a new visual assessment.

• needs_agent / blocked_pixel_target means a known obstacle blocks the approach

to the clicked surface or its stopping point. Inspect the new views and choose a safe alternative; do not repeat the same point or assume the target was reached.

• needs_agent / no_progress or exploration_failed means the hop ended without

sufficient approach or without reaching its exploration waypoint. Do not repeat the same image point or try to push through the same doorway. First inspect all returned views; move again only if they establish a clearly safe alternative. If glass, a closed/uncertain door, or a person prevents a safe route, remain stopped and report the physical blockage. An explicit operator/safety stop overrides exploration and blocked-audit requirements.

• For a find/search task, finish only when the target is clearly visible in a

fresh returned view, close enough for human-verifiable identity, and the requested location condition is satisfied. If it is only small or distant, approach via safe visible floor proxies while keeping a safety margin. Stop and cite the direction and visual evidence once close enough.

• In the normal path, do not call server_status, robot_protection_status, or

motion_safety_status; the authority proxy checks authority and battery on every movement call. Use diagnostics only after a refusal or abnormal result.

• Do not start a second hop while one is running. Use pixelgoal_stop on

operator stop or abnormal motion. reset_pixel_goal_map is operator-only.

Tool Invocation Discipline

• Call the navigation tool directly as pixelgoal_navigate on the dimos

MCP server. There is no namespace lookup step and no warmup or test call; never probe the tool surface with placeholder commands.

• The shell exists only to read skill files. Never use it for no-op or

placeholder commands (echo, true, printf, redirections to /dev/null), and never for navigation, perception, or note-taking. Every action is either a real dimos MCP tool call or a message to the operator.