version: "1.0.0" name: traverse-narrow-passages description: Navigate a BEHAVIOR v2 robot through an already-open doorway, narrow threshold, or tight aisle by safely staging outside the entrance, optionally pitching down only for a low obstacle that hides the floor overlay, squaring the chassis to the passage with adjust_chassis, then driving forward through the opening. After a successful forward, keep commanding forward until the trailing footprint is through; pitch, lateral translation, and side-on are only for a forward whose actual displacement falls far short of the request. Use only when the robot is already close to a passage or door and the opening's visible width is about the same as the blue floor path. Do not use for an ordinary corridor that is clearly wider than that path, to open doors, or to move obstacles.

Traverse narrow passages

<!-- codex-coordinate-contract:start --> The latest image in the conversation represents the current observed state. Older images are historical context; their points, bounding boxes, and object-part locations do not describe the current view. After any motion, re-locate the target in the latest image; do not shift, rotate, or scale an older location estimate to choose a new click.

Before EVERY point selection, inspect the latest attached clickable head image and re-locate the physical target. Use the full attached image as a relative coordinate canvas: integer u (column) and v (row) in 0..1000, with top-left (0,0) and bottom-right (1000,1000). Do not use encoded image pixels or crop coordinates. Bind every point, including every row of a multi-point call, to that exact current image_id. Never copy locations from older images, earlier reasoning, reference photos, tracking text, or overlay labels. Wrist images are observation-only and cannot be clicked. If the point is rejected, obtain a fresh head image and ground again; do not guess an image_id. <!-- codex-coordinate-contract:end -->

<!-- codex-task-lifecycle:start --> Follow this procedure only while traverse-narrow-passages is the active task Skill loaded by activate_skill. Its presence in history, including after compaction, does not make it active. After success and all required post-actions, return to the parent task; activate another matching Skill only when its procedure is needed. Instructions below to stay, retry, or exit describe pursuit of success and do not prohibit a change of approach or handoff to the parent task. <!-- codex-task-lifecycle:end -->

Follow the behavior-v2-baseline Skill and use only currently advertised behavior_v2 MCP tools. Never manipulate the door, door handle, frame, or a nearby object. Preserve the arm and gripper states. This Skill completes only when the complete trailing footprint has cleared the threshold. A stopped, stalled, timed-out, or ok=false sub-step requires reobservation or recovery. Do not end the session or exit this Skill.

Activate this Skill only after the robot is already close to a passage or door, and a fresh head view shows that opening's width is about the same as the blue floor path. A hallway or room entry that is clearly wider than the blue path is ordinary driving, not this Skill.

Required workflow

Follow these stages in order. Before considering side-on rotation, establish the designated staging pose immediately outside the entrance with the complete chassis still outside the threshold. Overlay conflicts beyond that endpoint are passage-fit evidence, not staging-route evidence; they cannot justify skipping the approach or choosing side-on from the initial distant observation.

1. Stage at the passage entrance

1. Begin with capture_head_camera. Identify the near threshold, both fixed

passage boundaries, any door leaf intruding into the opening, and a visible unobstructed floor point immediately before the entrance.

1. Call move_chassis_to_floor_point only when the selected point and the

complete direct translation to it stay on visible traversable floor and the complete chassis envelope has visible positive clearance from both passage boundaries and nearby obstacles along that finite route segment. The final chassis envelope must remain completely outside the threshold. Reconstruct the swept corridor only from the current pose through the preserved-yaw XY staging endpoint; do not extend the staging test through the passage. Do not substitute the blue base-forward overlay unless the direct displacement is forward-aligned. That overlay starts at the base front collision edge, is not a complete outline of the stationary chassis, and may continue far past the selected staging point. Ignore overlay conflicts strictly beyond the staging endpoint when deciding whether the finite approach is safe. Judge the whole finite reconstructed corridor, not only the clicked point or route centerline. This tool preserves yaw; the current yaw does not need to be suitable for crossing.

1. If the complete direct staging corridor is safe, the first chassis motion

must be move_chassis_to_floor_point. Do not treat merely being outside the passage, seeing the entrance, or having room to rotate as completion of staging. Skip this move only when fresh geometry proves that the complete chassis already occupies the designated pose immediately before the threshold.

1. If either boundary of the finite reconstructed staging corridor touches,

overlaps, or cannot be visibly separated from an obstacle before the chosen endpoint, do not use passage geometry beyond the endpoint to justify a side-on rotation at the current distant pose. Select another visible exterior staging point or approach only when its complete finite swept corridor is safe. If no safe route to an entrance staging pose can be established, do not enter the passage from this pose; choose another safe exterior staging route. Inability to reach staging is not evidence that side-on traversal is ready. Use a blue side rail as a staging boundary only for a forward-aligned finite displacement. If a staging attempt has already stalled, first obtain a fresh observation. Treat the exact inverse of the immediately preceding approach as a history-confirmed recovery path only when trusted tool history from the current physical episode identifies its last clear start pose, the fresh robot pose matches the recorded stalled end pose, the approach had no rotation and only one non-negligible translation axis, the scene has not changed, and the robot state is stable. First make one contact-relief adjust_chassis step of at most 0.05 m on that exact reverse axis, and no farther than necessary. Then command the measured remaining inverse path toward the recorded clear pose. Choose any intermediate endpoints from the recorded path geometry, current visible clearance, the advertised tool range, and a positive stopping margin; split the recovery only when uncertainty or tool feedback requires another check, not because of a universal per-call distance cap. Continue only when each result moves in the expected reverse direction and reduces the measured displacement to the clear pose. In every other case, retreat only when the reverse path is visibly safe. Never rotate while contact or separation is uncertain or before the complete chassis has returned outside the passage. If neither a safe recovery route nor a safe exterior inspection/rotation area is established, do not rotate or re-enter; seek a safe exterior recovery or inspection pose.

2. Expose the floor overlay

Only after reaching and freshly verifying the entrance staging pose. Staging pitch is optional and is only for seeing a hidden floor overlay. Do not pitch merely because this stage exists. Do not use staging pitch to decide stuck or side-on.

Pitch only when a low obstacle, such as a bed, sofa, or similar low furniture, hides the chassis or the floor overlay and a downward look is needed to see those. Then -50 is the usual adjust_pitch increment from the reset upright upper-body pose so the head camera can see the chassis. If the current head image already shows the chassis, do not pitch. adjust_pitch.degree is incremental: never stack another -50 after the chassis is already visible.

Do not pitch for a door, door leaf, door frame, or other tall vertical obstacle. Pitching the upper body into that geometry can jam the trunk or arms against the leaf or frame. Keep the upright pose and judge alignment from the current head image.

Do not pitch at a more distant exterior pose in order to choose the side-on branch.

3. Align the blue path outside the threshold

• Use adjust_chassis to rotate, translate left or right, and move forward or

backward until the complete blue path is aligned with the intended route. Its robot-base axes are forward for +X forward, positive translation for +Y left, and positive spin for counter-clockwise yaw. It has no leftward argument.

• Change only one of forward, translation, or spin per call throughout

this Skill. Choose the command magnitude from the measured center or yaw error, the complete verified swept path, the advertised tool range, and a positive stopping margin. Prefer a geometry-driven correction to arbitrary fixed-size increments; split a move only when visibility, clearance uncertainty, or tool feedback requires another observation. Do not impose universal 0.10 m translation or 10 degree rotation caps, and do not use tiny motions without a specific measured residual to correct. Inspect the model-visible post-action head image returned by each call; call capture_head_camera only when that image is missing or unusable.

• Keep the entire chassis outside the threshold while aligning. Align both the

near and far blue path edges through the narrowest part of the passage, not merely its centerline.

• Separate center/yaw misalignment from a width deficit. Rotation and lateral

translation can place the blue path between the passage boundaries, but they cannot make the blue path narrower. Align by putting as much of the blue path as possible inside the visible passage floor. Unused passage floor on one side of the blue path or chassis is a center offset: translate toward that floor. Positive translation is left; negative translation is right. A blue rail or chassis edge on the bed, wall, or other boundary while the opposite side still shows unused passage floor is unfinished alignment, not a width deficit. Use lateral translation when the opening center is visibly offset from the blue-path center and the opposite blue edge still has clearance. Never translate repeatedly toward one boundary merely to clear the other blue edge.

• Establish yaw from directions, not from raw gap changes. For a forward route,

compare the blue longitudinal rails with the fixed passage's longitudinal direction across at least two separated slices. A yaw error exists when those directions are not parallel. Only after directional alignment, derive center offset from the displacement between the blue route centerline and the local midpoint of the two passage boundaries at corresponding slices. A tapered, stepped, or locally narrower passage can change raw left-right gaps without a yaw error; classify that geometry separately as a local center shift, width deficit, or obstacle. If the longitudinal direction or local boundary pair is not observable, stop that yaw inference instead of inferring yaw from gap imbalance; obtain a clearer observation.

• After each alignment motion, identify the limiting blue edge and passage

boundary in the returned image. Do not issue another translation in the same direction unless the previous move visibly reduced a measured center offset without creating or worsening a conflict at the opposite edge. Use no fixed call-count limit: continue only while each measured result has the expected sign, reduces the quantified residual center offset, and adds usable post-action evidence. Stop on the first result that fails any of those conditions. A tight or touching blue rail is alignment or overlay geometry, not permission to rotate side-on. If the passage direction or both fixed sides are still ambiguous, stop that heading guess and obtain a clearer observation instead of inventing a side-on heading.

4. Square on, then keep forwarding until through

From the verified entrance staging pose, use adjust_chassis until the chassis faces the passage squarely. Square-on means the blue longitudinal rails are visually parallel to the passage direction across at least two slices, and the blue-path center sits on the local midpoint of the two fixed sides. Use spin for yaw and translation only for a measured center offset.

Once this square-on alignment is established, alignment is finished. Do not issue another translation, extra retreat, or spin to refine center, enlarge a rotation circle, or reach an open-area spin pose. The next chassis command must be forward through the passage.

Then command forward. After crossing starts, a successful forward means the chassis is not stuck. Compare requested with actual. Treat the step as successful when actual.forward_m has the requested sign and the commanded distance was completed (linear_target_reached or near_target_ok, or |actual| is not far below |requested|). After that success, command only more forward until the complete trailing footprint has cleared the threshold. Do not translation, spin, or adjust_pitch after a successful forward. Unused floor, a visible door leaf, or a blue-rail gap in the image does not authorize adjustment after a successful forward.

A forward is stuck enough to stop that attempt only when the actual displacement falls far short of the request: a large |requested|-|actual|, |actual_axis| < 0.20 * |requested_axis| when |requested_axis| >= 0.05 m, obstacle_limited without reaching, or a velocity stall with almost no progress. Do not keep forwarding in that same attempt, and do not treat leftover floor beside the path as proof that the last forward succeeded.

Only after that shortfall: do not translate, retreat farther for rotation room, or keep aligning. If any part of the chassis is between the passage boundaries, one reverse forward on the same axis is allowed solely to put the complete chassis outside the threshold. Then go to Stage 5 and spin in place. Do not start side-on from a successful forward or from a predicted tight fit.

Treat the blue overlay as a base-forward swept-path corridor that begins at the base front collision edge, not as a complete outline of the stationary chassis. For a forward-aligned motion, use its rails, fill, and scale as heading and scale evidence while squaring on and choosing forward size. A completed forward step that made progress is valid only when a visible positive gap separates each aligned blue rail from its passage boundary at the same narrowest longitudinal slice and the unrendered current chassis envelope also remains clear. A blue rail that coincides with, touches, or cannot be visibly separated from a post, door leaf, wall, bed, or other obstacle is a width deficit in the overlay, not a completed forward fit and not by itself a side-on decision. Side-on waits until that real forward has already fallen far short. Do not try to make the blue path narrower with further lateral translation once alignment is finished.

5. Use an image-aligned side-on crossing only when needed

Use side-on traversal only after Stage 4: a real forward from the aligned entrance fell far short of the request. Do not rotate side-on from a predicted tight fit or after a successful forward. The chassis must still be completely outside the threshold. Never rotate while any part of the chassis is between the passage boundaries.

The chassis is wider left-right than front-back. Side-on means the long left-right axis lies along the passage, so the short front-back axis spans the gap. In the current head image, the blue overlay's near/lower transverse border is the nearest full-width line at the base front collision edge, marked by the 0.2 m and 0.4 m lateral labels and the short forward ticks on that line. That line is the chassis left-right axis now. Do not confuse it with a longitudinal blue side rail, route centerline, direction arrow, or 0.5 m forward-distance tick.

1. Choose the turn direction from this image by which way is more

convenient for the later crossing and the work after it. Positive spin is counter-clockwise (front swings toward chassis-left). Negative spin is clockwise (front swings toward chassis-right).

1. Spin in place at the aligned pose. If Stage 4 did square the

chassis to the entrance, command about 90 in that chosen sign. Do not translation or extra forward before this spin to create rotation-sweep clearance. The post-exit rotation-sweep proof below does not apply to this Stage 5 spin.

1. That in-place spin is the last yaw change. Do not add another

spin to chase residual angle or a prettier lower-border line. Requested spin and actual.spin_deg only confirm that rotation happened.

1. After this spin, the chassis center is still the Stage 4 aligned

pose: that is the most accurate occupancy of the passage. Do not forward, extra translation to recenter, or retreat to "sit in the strip." The next chassis command must be the signed translation that crosses along the passage.

1. Apply the signed-translation rule below and cross. After a correct

side-on yaw, translation moves the long left-right axis along the passage; the blue forward overlay now points across the gap and is not the crossing.

Before rotating, preserve a stable exterior image, base pose, yaw, the near and far threshold references, both fixed boundaries, the intended signed lateral motion axis, and the complete planned swept corridor from the current footprint through the far-side exit. The later side-on head view cannot replace this record. After rotation, the blue forward overlay remains chassis-forward and points across the passage, not along the signed translation; it is not the crossing and cannot prove lateral-sweep clearance. Neither can a stationary footprint gap or an overlay ending at the camera-facing wall.

After the image shows the blue lower border parallel to the passage, separately reconstruct the stationary chassis envelope and its signed lateral sweep; require visible positive gaps on both opposing sides at the near threshold, narrowest slice, and farther swept corridor. Do not infer those gaps from the blue longitudinal rails or fill after rotation, because they still describe base-forward motion across the passage. Do not accept a local stationary gap when the reconstructed chassis sweep converges toward a boundary farther through the passage. If the lower border, passage direction, chassis envelope, or any limiting clearance is ambiguous, do not rotate on that guessed geometry; obtain a clearer observation. Stop rotating as soon as these image and corridor conditions hold.

If a nonzero spin returns actual.spin_deg=0, the command produced no measured rotation even if top-level ok=true; treat it as a controller dead-zone/no-motion result, not a transport failure and not progress. Never repeat that command or issue another small corrective spin to chase the residual angle. If the image already proves that the blue lower border is parallel to the passage direction and the independently reconstructed lateral chassis corridor has positive clearance, keep yaw fixed and start the signed translation immediately. Otherwise abandon this side-on alignment attempt; return to a verified exterior staging pose and choose another route. Do not let a nominal yaw target override the missing image evidence. Do not insert a post-spin forward to "occupy" the passage.

Transform the preserved passage direction through the measured actual pose to select the signed base translation axis; do not infer the sign solely from a nominal rotation amount. When the desired crossing direction maps to current chassis-right, use negative translation; when it maps to current chassis-left, use positive translation. A completed spin is not itself a crossing command; the first chassis command after that spin is this signed translation. The pre-spin square-on pose already occupies the intended passage; do not require a post-spin forward occupancy step. The blue forward overlay after rotation cannot supply lateral-sweep proof. If the signed corridor cannot be reconstructed from the preserved image sequence, scale, and actual pose transform, reconstruct it from a clearer observation without leaving the aligned side-on pose. Once the in-place spin and the signed-translation axis are established, keep yaw fixed and choose each single-axis translation endpoint from that verified corridor and the remaining distance needed to clear the trailing chassis envelope.

Cross and clear the doorway

For a valid forward route, choose each forward endpoint from the visible clear corridor, scale, required remaining displacement, and a positive stopping margin. For a side-on route, use the signed translation established above. After every move, verify that the route remains unobstructed and the newest robot state shows no unexpected chassis height, tilt, collision, yaw drift, or reverse progress. Halt the current chassis or arm command on collision evidence or unstable chassis state, then recover from the observed state.

Treat top-level ok=true as confirmation that the tool call completed, not that the requested chassis displacement succeeded. After every adjust_chassis call, compare requested with actual, inspect linear_target_reached when it is reported, and inspect obstacle_limited and obstacle_stop_reason. Count only measured actual displacement along the commanded axis as progress. Let requested_axis and actual_axis be the signed values for that one commanded linear axis. When |requested_axis| >= 0.05 m, require actual_axis to have the same expected sign and |actual_axis| >= 0.20 * |requested_axis|. Treat zero, opposite-sign motion, or |actual_axis| < 0.20 * |requested_axis| as a hard stall even if top-level ok=true. Apply this symmetrically to positive and negative forward or translation requests. Thus, a +0.10 m or -0.10 m request that moves only 0.0037 m, 0.0007 m, or 0.0009 m in magnitude is obviously blocked; the first such result must end that motion attempt. If obstacle_limited=true, linear_target_reached=false, a stall is reported, or this below-20% rule applies, stop immediately and never repeat the same motion direction. This ends that motion attempt, not this Skill or the session. After a successful forward, those flags are not a reason to stop and adjust; keep commanding forward. After a large requested-versus-actual shortfall, do not keep forwarding and do not use leftover floor as permission to realign. Do not inspect-pitch or translation to recenter; go to Stage 5 and spin in place. Treat a reported velocity stall, partial target failure, or uncommanded yaw drift as decisive even when obstacle_limited=false and no dynamic collision is reported. Do not add millimeter-scale results across retries or claim progress from them. Do not claim cumulative progress from requested distances.

The history-confirmed exact-reverse recovery in Stage 1 applies only after a hard stall with no unused opposite-side passage floor left to absorb a center correction. It reverses rather than retries the blocked direction and applies to a true-contact stall during either staging or crossing. For a side-on crossing stall, first obtain a fresh observation and require its pose to match the recorded stalled end in the same unchanged episode. Sum the measured signed translations since the last trusted pose whose complete footprint was still outside the passage. Use one necessary contact-relief reverse move of at most 0.05 m, then reduce the measured remaining displacement along the same axis toward that fully exterior pose. A successful intermediate slice inside the passage is not a trusted exterior recovery endpoint. Stop that recovery on any reverse stall, wrong-way displacement, unexpected yaw or lateral drift, unstable chassis state, or new obstacle evidence. Do not rotate until a fresh observation confirms that the complete chassis is back outside the passage and separated from both boundaries.

After recovery, never retry the stalled centerline. A different centerline is eligible only when new observations identify the fixed boundary that limited each previous attempt and the transformed swept-corridor proof shows positive clearance at every previously stalled longitudinal slice. Reclassify the stall from geometry, not raw near-to-far gap trend: compare the route direction with the fixed passage direction, then compare the direction-aligned route centerline with each slice's local passage midpoint and width. For a side-on route, use the near/lower transverse border as the direction reference; for a forward route, use the blue longitudinal rails. Treat taper, a step, or a local bottleneck as passage geometry rather than automatic yaw evidence. Retaining the same unverified yaw while translating to another centerline is not a different route. If the route direction diverged from the passage direction, correct yaw while fully outside and re-establish the applicable visual parallelism before considering a new swept corridor. Merely shifting the base, seeing local gaps at the new stationary pose, or guessing the midpoint between stalled lines is insufficient. Repeated stalls do not authorize blind probing or establish a useful centerline by interpolation. If the limiting boundaries or any stalled slice remain unobservable, do not claim traversal success; seek a newly observable route or a different proven corridor.

Use the observation sequence, scale overlay, and cumulative chassis displacement to clear the trailing footprint, rather than stopping as soon as the camera or leading edge crosses the threshold. Continue only while the far-side route is visibly clear.

Before crossing, preserve the near and far threshold references and the gap from the leading side of the footprint to the near threshold in the last verified exterior observation. For a side-on route, use the complete near-edge lateral scale, or reported path_width_m, only as a conservative upper bound on the chassis-Y motion-axis span needed to clear the trailing side. It is not an exact stationary-footprint measurement or cross-passage clearance proof. Do not mistake a half-width marker for the full span. Accumulate only actual.translation_m values with the expected sign. Its magnitude must cover the leading-side-to-near-threshold gap, the threshold-band depth, the full motion-axis footprint span, and a positive far-side margin. Do not treat one footprint span from the starting base pose as enough. If one of those metric terms is visually uncertain, use the sequence of the same fixed threshold references to establish that they receded past the trailing side; otherwise leave traversal unconfirmed and keep using the image sequence; do not invent a distance.

Use the latest usable post-action image as the fresh final observation. Call capture_head_camera only when that image is missing or ambiguous, or when a stationary observation is needed to establish settling or clearance. For a forward route, require open traversable floor in the forward direction and no passage boundary in the forward exit corridor. For a side-on route, do not require open traversable floor in the robot's current forward direction: that axis faces across the passage, so a fixed side boundary may occlude the projected forward blue path beyond its near edge. Instead require the latest observation and image sequence to show positive separation between the stationary footprint and the side boundaries, a clear signed lateral exit corridor, and the fixed threshold references behind the trailing side according to the metric bound above. Do not confuse an occluded forward path projection with a blocked signed lateral exit corridor, but treat any contact or ambiguous separation at the robot origin as unconfirmed.

After the complete trailing footprint has cleared, do not rotate merely to inspect the room, face a downstream target, or begin another Skill until the entire rotation sweep is proven clear for the chassis and the preserved arms and grippers. This post-exit rule does not apply to the Stage 5 side-on spin and must not delay or displace that in-place spin. The blue overlay represents only the base-forward swept corridor; it does not prove clearance for an arm, wrist, finger, or gripper housing that sweeps toward a wall, frame, table, or other obstacle. A single depth point on a panel in front of the base, or clearance on only the camera-facing side, is not a swept-volume proof. Establish positive clearance on every side used by the planned rotation, using fresh head and wrist evidence when the head view does not expose the appendage-side gap. If that complete envelope is not observable, keep the verified exit yaw and stop instead of rotating to search.

After every post-exit rotation, compare the preserved and current arm joint and gripper states as well as the returned head or wrist images. Treat uncommanded arm-joint displacement or velocity, a new left-right finger-opening asymmetry, close obstacle geometry inside a wrist grasp volume, or loss of visible separation from a fixed boundary as contact evidence even when the chassis tool reports no dynamic collision or commanded translation. Stop all chassis and arm motion immediately; do not retry the same rotation direction, extend an arm, or start a grasp. Do not reverse the rotation unless a fresh observation proves that the complete reverse sweep back to the last contact-free pose is clear; otherwise hold the contact-risk pose and keep observing until a contact-free recovery is supported. Hand off to a manipulation Skill only after a stable, contact-free post-exit pose is verified and the intended target is freshly and uniquely reacquired. This is a parent-task Skill switch, not session end.

Do not require the head camera to see the whole chassis; it cannot provide that view. If trailing-footprint clearance cannot be established from the scale and motion history, do not claim success. Report the final image_id and concrete visual and state evidence only when claiming a completed crossing. After successful crossing and any required post-exit checks, return to the parent task or activate a manipulation Skill only when its procedure is needed.