zju3dv/supernav/global-navigation-pixelgoal

Global Navigation (Go2 PixelGoal)

These rules guide global navigation through the Go2 PixelGoal tools. Use MUST, MUST NOT, SHOULD, and MAY literally. Rule IDs remain stable audit labels.

Load the Needed Rules

• Before leaving the local area, taking a long return, or committing to a long hop,

read references/exploration-and-return.md completely.

• When the target or its room is visible only through glass, or no passable entrance is

visible near the start, read references/entrance-search.md completely before searching the perimeter or reporting blocked.

• At the first stall, or before reporting blocked, read

references/recovery-and-stop.md completely.

Go2 Contract

• NM-LIFE-001 — One initialization. The harness supplies one initial fresh

front/right/back/left surround. MUST NOT call observation tools just to repeat it.

• NM-LIFE-002 — PixelGoal rules. MUST read the complete run-local

go2-pixelgoal/SKILL.md before the first pixelgoal_navigate. That skill owns safe point selection and result handling; this skill owns global search decisions. Its permitted surfaces include traversable floor and stable vertical surfaces under the conditions it defines. Apply those rules throughout this skill and its references; route selection does not impose a floor-only point restriction.

• NM-LIFE-003 — Evidence boundary. MUST perceive and act only through the

whitelisted DimOS MCP tools and their latest returned images. MUST NOT use shell, Python, files, artifacts, maps, coordinates, trajectories, or hidden robot state for navigation decisions.

• NM-LIFE-004 — Formal finish. There is no session-close tool. MUST answer only

after the achieved gate holds on the latest surround, or after writing the complete blocked audit required by the recovery reference. On operator stop or abnormal motion, MUST call pixelgoal_stop before answering.

Target and Observation Gate

• NM-OBS-001 — Complete target hypothesis. MUST keep a compact hypothesis containing

the requested object/region, distinguishing appearance or material, required relations, room context, and final visual evidence. A category match alone is insufficient.

• NM-OBS-002 — Observe-decide-act. After initialization and every hop, MUST inspect

front, right, back, and left, including both horizontal edges; test the full target hypothesis; census every visible transition; choose one bounded action; and repeat on the returned surround. Every completed pixelgoal_navigate, including a needs_agent handoff, returns fresh available views. exploration_complete calls for a new census, not success or blocked reporting. Use observe_surround only for missing/stale-image recovery.

• NM-OBS-003 — Census before decision. MUST classify each doorway, side opening,

corridor continuation, and room interior as traversed, meaningfully inspected, or new/uncertain. MUST list new/uncertain choices before selecting a proxy using:

census: front=...; right=...; back=...; left=... | new=... | frontier=... | next=... | why=...

Current pixels outrank earlier notes. An empty noted list is not exhaustion. A hop that empties the list may expose the next junction. MUST NOT report blocked or begin a return in the same decision as a successful hop; first census its fresh surround.

Core Navigation Decisions

• NM-MOVE-001 — Relevance before novelty. SHOULD prefer current target evidence or a

target-relevant current route, then a safe unfamiliar opening, a current unentered sibling, and finally a bounded return to one exact older noted opening. A clearly mismatched room MAY be deferred while a stronger route exists, but MUST stay noted.

• NM-MOVE-002 — Visible bounded movement. All locomotion is one

pixelgoal_navigate(point=[x, y], direction=...) call on a safe point selected according to go2-pixelgoal in the latest view. Reaching a proxy is not completion. Every hop or refresh expires old image points.

• NM-MOVE-003 — Hop gate. SHOULD choose the farthest safe point allowed by

go2-pixelgoal that advances through the selected transition without skipping a more relevant local choice or an unrecorded opening. Otherwise use an intermediate visible proxy. Read the exploration reference before leaving the local area or committing to a long return.

• NM-MOVE-004 — Information gain when nothing matches. When no current view shows

target evidence, MUST place the point where the hop buys the most new information: through the most promising new/uncertain census opening, far enough past the threshold that the next surround reveals new area. MUST NOT spend a hop re-showing the same room while an unentered opening is visible.

Real-World Adaptation (0.6 m Go2 rig)

• The cameras sit about 0.6 m above the floor. Chair legs, cables, thresholds,

and bag straps are real collisions even when they look small; give them margin. You cannot see over desks, shelves, or cubicle walls — a blocked high view does not mean the floor is blocked; judge floor from the lower half of every view.

• Offices have glass partitions and polished floors. Glass shows unreachable

space (follow entrance-search); floor glare and reflections are not obstacles. Dark glass can look like open space — confirm with the other three views before marking near it.

• Crossing a doorway: prefer a point clearly past the threshold on the far-side

floor. Apply go2-pixelgoal to any alternative surface; never mark the threshold strip itself, and confirm the opening is passable.

• People may walk through the scene; if legs or moving objects cover the

intended point, wait one observation cycle instead of marking near them.

Stop Gates

• NM-TERM-001 — Achieved. MUST finish achieved only when the latest fresh surround

shows the complete target, distinguishing attributes, required relations, and matching context at human-verifiable distance. Reaching a waypoint is never success by itself.

• NM-TERM-002 — Blocked. MUST read the recovery reference and exhaust its bounded gate

before reporting blocked. MUST write its blocked audit with every clause true before answering.