version: "1.0.0" name: place-object-in-container description: Place an object already held by the robot into a specified open container and visually verify that the object is inside. Use for requests such as place object in container, put the held item into a bin, box, basket, or other open receptacle, or for the placement stage of a pick-and-place task. Do not use for picking or regrasping an object that is not currently held.

Place object in container

<!-- 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 place-object-in-container 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. Codex has no explicit lifecycle-close action: a Skill handoff is expressed by returning to the parent task, not by inventing a tool call. Before every handoff, stow any arm that is no longer needed as required below. 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 -->

Use the advertised behavior_v2 tools and the baseline Skill. Treat this as a state-driven placement task, not a fixed action script. Prefer direct measured progress over ritual observation, fixed step schedules, retry counters, or controller status labels. The only Skill exit is a fresh observation that the intended object has settled inside the intended container. Stopping a current motion, bind, or descent is a sub-step change, not Skill exit and not session end.

Necessary hard gates

These are the only hard gates in this Skill:

• Establish which arm holds the intended object and identify the target

container before moving or releasing. Use the freshest reliable evidence available, including the immediately preceding pick result, current head or wrist images, and grasp-volume evidence. Do not force a left-and-right wrist capture pair when the holding arm is already unambiguous. If the holding state, object identity, or target container cannot yet be established from obtainable current evidence, do not release; keep observing and re-establish identity from new evidence. Stay in this Skill.

• Keep the holding gripper closed until the release conditions below are met.

Do not move or open the non-holding gripper as part of placement.

• Before a coordinate-based motion, prefer object and container geometry from

the latest persistent_tracking or a fresh bind. If that geometry is untrustworthy, re-bind. Coordinates recited more than about 10 turns ago are probably no longer applicable. If usable current geometry is not yet available, do not release; keep observing or re-bind. Stay in this Skill.

• If the placement point is far away, use adjust_chassis, adjust_height,

or adjust_pitch to close the gap by hand, or use move_to_reach_point to close the distance automatically. If you choose move_to_reach_point, call it only after spin_to_facing_point and the post-facing blue path is clear. move_to_reach_point is a straight-line chassis move toward the selected point. It does not weave around furniture. Any obstacle on that line will stall or jam the robot. If a bed, chair, table, wall, or other object blocks that line, first go around with adjust_chassis or move_chassis_to_floor_point. Do not call move_to_reach_point until the robot has faced the point and the blue path is clear. After that detour, face again, then approach with move_to_reach_point before any planar EEF placement.

• Stop the current motion direction on actual collision or contact, loss of

clearance, loss of the grasp, or measured motion that is unsafe. If no evidence-supported safe recovery path is currently available, do not release; change viewpoint or geometry and keep looking for a safe recovery. Stay in this Skill.

• After each interior bind, state interior_surface=far_wall or

interior_surface=floor before any planar holding-arm XY or descent. The only two legal interior surfaces are the container floor and the far inner wall. For an ashcan or trash can, do not use far_wall. Empty-ashcan interior must be the visible bottom disk with Z below 0.01 m, or an already-placed in-can soda (no Z gate). If empty-bottom interior Z is greater than 0.01 m, that click is the side wall: re-bind the bottom. Do not start planar XY or release on that wall click. If the quadrilateral XY branch below is active, also state align_mode=constraint_rect and the current x_lo, x_hi, y_lo, and y_hi. Release only when the classified XY band below is satisfied and current visual evidence shows that the object's complete downward footprint is inside the opening, the fall path and rim clearance are safe, and the release height is appropriate for the visible geometry. When align_mode=constraint_rect is stated, that rectangle inequality is the classified XY band. For an elongated object such as a board, puzzle sheet, or box, both visible lengthwise endpoints of that downward projection and the gravity-aligned sweep volume must fit inside the opening. A grip point, centroid, visible narrow edge, tracked interior point, or any single pixel on the object does not authorize release. A large-looking opening does not authorize release while the classified XY band is missed.

• Report success only after a fresh observation shows the intended object

settled inside the intended container, and set_arm_to_grasp_position has returned both arms to the grasp-prep pose with gripper="open". That report is this Skill's only exit. It does not end the Codex session. Return to the parent task; if more requested objects remain, activate the next matching Skill.

• Before every Skill exit (success, cancel, or handoff to parent Second

logic or pick-up-object), if the object-holding arm is no longer needed, call set_arm_to_grasp_position with gripper="open" on that arm. arm="both" is required. Do not start move_to_reach_point or spin the chassis with an outstretched unused arm. A small withdrawal to unocclude is not a stow. Keep this habit: when an arm's job is done, set-arm it and leave the workspace clean.

• If the object falls from the gripper, this Skill cannot finish placement

while empty-handed. First stow any unused placement arm with set_arm_to_grasp_position as required above, then return to the parent task. The parent chooses the next unused can. A can left outside after a miss is not that next can unless the parent last-step far-can rule says so. Do not immediately re-pick this same object inside this Skill. Do not report placement success. Do not end the session from this Skill; after returning, the parent may end the session if its last-step check says every remaining outside can is under 0.5 m from the ashcan. Do not attempt an improvised regrasp in this Skill.

• If the ashcan interior is not visible, this Skill cannot finish

placement. First stow any unused placement arm with set_arm_to_grasp_position as required above, then return to the parent task so the parent Second logic can stand the ashcan, then resume the parent pick/place loop. Do not improvise a stand-up procedure in this Skill.

Do not add hard action counts, alignment cycles, retry budgets, fixed orientation tolerances, or fixed IK-residual thresholds. Do not add universal descent-distance schedules. The 30 cm near_rim clearance below is this Skill's hard height gate while the classified XY band is not yet satisfied: the held object and the holding gripper/wrist must stay at least 30 cm above current near_rim Z and must never go below that. After the classified XY band is satisfied, the lower bound becomes 0.20 m: descend into the 0.20–0.25 m held_z - near_rim_z band. After that descent, small XY corrections stay at the current height; do not lift back to the 30 cm planar-travel gate. Visually check, then release. The classified XY bands in Align are this Skill's planar-target gates, not optional hints. A tool's ok, timeout, non-convergence, final position error, or ori_err_deg is diagnostic information, not by itself a reason to stop. In particular, EEF orientation error must never override safe measured XYZ progress, current containment, grasp stability, or collision evidence.

Establish current state

Determine the holding arm without replaying preparation actions that have already completed. Wrist images are useful when current holding evidence is ambiguous: red overlay pixels are visible non-robot 3D points inside that gripper's grasp volume, and they support a grasp only when they overlap the intended object's visible surface. Near-field occupancy may identify which gripper is occupied, but it does not prove object identity. Obtain additional head or wrist evidence only when needed to resolve a material ambiguity.

Do not call set_arm_to_grasp_position merely because placement started. Use it when the current arm pose itself prevents a safe observation or reach, including the one-shot out-of-view recovery below and the stow-before-reach fallback in this Skill, and preserve the grasp. After that arm's job is done, the exit stow above is required, not optional. Prefer a model-visible post-action head image when it is current and usable. Call capture_head_camera only when the latest result has no usable current image or the scene has changed enough to make it unreliable. Do not capture after every action by default.

Use visible image points for runtime geometry. BDDL/task labels such as ashcan 1 describe the goal but are not guaranteed runtime object names.

Use persistent tracking

This Skill does not schedule when to bind or re-bind. The two XY branches below say what a placement bind contains, not when to call track_object_distance. Bind or re-bind whenever current geometry is missing or untrustworthy. If you suspect a click missed the intended surface or that a bind is stale, look at a fresh clickable head image and re-click the visible target. Do not bury yourself reconciling historical coordinates, blending old and new tracks, or computing a click from prior images; look at the current image.

Inspect persistent_tracking on the latest tool result. track_object_distance is an atomic full-set replacement, not an append or read operation: a successful call replaces all previous bindings. Once a point is bound, the tracker keeps following that named surface in the live head RGB-D for as long as the point remains in view. persistent_tracking after each action is the only current measurement. After initialization, consume the updated coordinates from persistent_tracking after every action. observation_sequence, current u/v, and xyz_in_robot_base_coord_m describe the current update. source_image_id, source_session_id, and source_observation_sequence are seed provenance only.

Coordinates recited more than about 10 turns ago are probably no longer applicable, especially after the robot has moved. Prefer the latest persistent_tracking when reading object and container geometry. If that snapshot is untrustworthy, re-bind.

Both XY branches include:

• a stable visible patch on the held object
• a near_rim point on the visible rim closest to the holding arm and the

planned travel path

near_rim is the measured height of that approaching edge. It is required for this geometry, not an optional extra. Additional rim or corner anchors beyond that approaching-edge point remain optional. The remaining interior points belong to the chosen XY branch below. Do not use an interior Z as rim height. Select broad visible material patches. Do not select background, an image boundary, a depth discontinuity, the rim as the interior, or a guessed continuation of an occluded surface. The best placement viewpoint shows the container floor, or a can already sitting inside the opening. Get that view before trusting an interior bind. If the empty ashcan bottom is not visible, lean the trunk forward with one adjust_pitch of -10 degrees, then look again. Do not command a larger pitch. Do not keep pitching once the bottom disk or an in-can soda is visible. After each lean, re-bind and keep clearance_z at least 0.30 m. Raise, lower, or step the chassis only if that one lean is not enough. Bind interior on that floor patch, with margin from every opening edge, or on the in-can object. Those surfaces are much more accurate than a side wall. The dark disk in a steep top-down ashcan opening is usually the inner funnel wall, not the floor. Tracking a side wall as interior often drives the held object into the trash can or dumps it outside. Do not start planar XY or release on a wall click when a floor or in-can bind is obtainable. interior_surface=floor only when the click is the container bottom or an object resting there, with Z down on that bottom — not merely below near_rim. A mid-bin Z is a wall. Empty-ashcan floor bind: interior Z must be below 0.01 m. If interior Z is greater than 0.01 m, the click is the side wall, not the floor. Re-bind on the visible bottom disk. Do not start planar XY or release on that wall click. An already-placed in-can soda as interior has no Z gate. near_rim Z must be greater than 0.2 m. That is rim height, not clearance_z. If near_rim Z is not greater than 0.2 m, re-bind on the rim band. If it is below 0.01 m, the click is the room floor. Do not use that floor click as rim height. If a previously placed soda is already inside the ashcan, bind interior on that visible in-can soda; do not click the empty funnel wall or a point outside the opening. If the only usable empty-ashcan floor click sits near a boundary, still require Z below 0.01 m and re-bind a more centered bottom disk. Do not switch that click to the far_wall Align branch.

A bind UV is a model annotation and can miss the intended surface. After a bind, if the holding EEF moves but that object's track (u/v and xyz_in_robot_base_coord_m) stays essentially unchanged, that mismatch is an annotation failure: the click did not land on the held object. That geometry is untrustworthy; re-bind the object point on a broad visible patch of the actual held object, keeping every other still-needed point in the same replacement call. Do not keep moving the EEF as if the unchanged track were the object.

If a placement motion accidentally carries the held object out of the current head view, or the held object is not visible, call set_arm_to_grasp_position once for the holding arm with gripper="keep" so the object returns to a visible stow pose. Then bring the held object into the head image: left holding arm adjust_left_eef_pose_in_head_frame with leftward=-0.1; right holding arm adjust_right_eef_pose_in_head_frame with leftward=0.1. Then re-bind track_object_distance on that new image. Do not repeat set-arm as a search. This is a sub-step recovery, not Skill exit.

Because replacement is atomic, include every point that must remain active in the replacement call. Treat tracker confidence and status="observed" as measurement quality, not semantic identity proof. If coordinates and current visual evidence materially contradict each other, the geometry is untrustworthy.

Approach the container

Approach and placement are different stages. Do not apply the placement-stage arm-XY-first rule while the container is still out of reach. If the placement point is far away, use adjust_chassis, adjust_height, or adjust_pitch to close the gap by hand, or use move_to_reach_point to close the distance automatically.

move_to_reach_point is a straight-line chassis move toward the selected image point. It keeps the current yaw and drives the shortest line to a reach pose. It does not plan around obstacles. If a bed, chair, table, wall, or any other object sits on that line, the command will stall or jam. A visible container does not authorize the call. The blue path overlay is the forward corridor for the current yaw. If the selected point is still off to one side, that overlay is not the route the reach will drive; only after spin_to_facing_point is the blue path the actual travel route. Before every move_to_reach_point, call spin_to_facing_point with the current head image_id and target u/v, reidentify the same target on the post-facing image, and call the reach only when that blue path is clear for the full chassis footprint from here to that reach pose. Do not use spin_to_facing_point as a substitute for the reach; facing only sets yaw.

If the target container is visible in the current head image and the holding arm cannot yet place over the opening, but the blue path is blocked, do not call move_to_reach_point yet. First go around the obstacle until a clear straight corridor exists:

• Use move_chassis_to_floor_point to a visible traversable floor point on

the clear side of the obstacle, requiring the complete path and final chassis footprint to stay clear.

• Use adjust_chassis for a measured forward, backward, lateral, or spin

correction that increases clearance and opens that straight corridor.

• Use traverse-narrow-passages only when a narrow legal passage itself must

be crossed.

Those chassis tools are the detour, not the final reach. After the new image shows a clear straight corridor, call spin_to_facing_point on that current head image_id and a visible point on the container, inspect the blue path, and call move_to_reach_point only when that overlay is clear. That call is the reach onto the opening. A search turn is allowed to bring the container into view; seeing the container is not enough if the blue path is still blocked.

Prefer a stable visible point on the opening interior, the rim, or a nearby staging point on the same container body. Inspect the actual move_to_reach_point completion evidence and the post-action image. Never reuse the pre-motion UV for later placement.

If the selected point is on the floor, a depth discontinuity, the wrong surface, or otherwise contradicts the visible container, that geometry is untrustworthy; re-bind a nearby visible point on the same container. Do not treat a bad click as a reason to skip the reach. Retry the reach after the straight path is still clear. Stay in this Skill.

If move_to_reach_point stalls on contact, treat that as a blocked straight path, not Skill exit and not session end. Stop the current direction, back out to a clear footprint, go around with adjust_chassis or move_chassis_to_floor_point, then spin_to_facing_point again, and call move_to_reach_point only after a new image shows a clear blue path.

After a successful reach, or after chassis / height / pitch adjustment has put the container at a clearly reachable place pose, consume the post-action image. The two XY branches below describe the placement bind; when to bind or re-bind remains the model's decision.

Align from current coordinates

This section applies only after the container is at a clearly reachable pose for the holding arm. Arm XY first, with optional whole-body reach, is a placement-stage rule. It is not the approach rule.

The robot-base axes are X forward, Y left, and Z up. For adjust_*_eef_pose_in_head_frame, the x, y, and z arguments are robot-base deltas. The forward, leftward, and upward arguments are in the starting head-camera frame. Do not mix these parameter families in one call.

Choose the XY branch

Judge the container floor by what the container is, not by whether the current view looks narrow or occluded:

• Small floor: an ash can, trash can, basin, or bowl. Use the

single-point XY branch.

• Large floor: a toolbox, toolkit, or drawer. Use the quadrilateral XY

branch. A divider or one visible bay does not make it small.

If the container is not in those lists, treat a trash-can-like or bowl-like opening as small, and a toolbox-like or drawer-like opening as large. State floor_size=small or floor_size=large and the chosen branch. Both branches share the held-object bind, near_rim, rim clearance, and travel rules in this section. They differ only in the XY target.

After every interior bind or re-bind, and before any planar holding-arm XY or descent, output one line that names the bound interior surface. Prefer the current bind or the latest persistent_tracking for x_lo, x_hi, y_lo, y_hi, interior Z, and near_rim Z:

interior_surface=far_wall

or

interior_surface=floor

Include the current interior Z and near_rim Z in that same message. far_wall means the far inner wall, the wall farther from the robot along +X, with Z near the rim and not on the floor. Do not use far_wall on an ashcan. floor means the container floor inside the opening. For an empty ashcan that requires interior Z below 0.01 m, not merely below near_rim. A mid-bin Z above 0.01 m is the side wall. An in-can soda as interior has no Z gate. Those are the only two interior surfaces the model can bind. The quadrilateral branch is an XY rule on floor; it does not add a third interior surface. If that branch is active, include align_mode=constraint_rect and the current x_lo, x_hi, y_lo, and y_hi in the same message. If a click is not clearly floor or far_wall, that geometry is untrustworthy. Do not start planar XY or descent until that line has been stated for the current bind.

Clear the approaching rim

Current near_rim Z is the edge-height gate for the first planar approach, not a reason to climb back after descent. Compute clearance_z = held_z - near_rim_z. While the classified XY band is not yet satisfied, never go below 0.30 m. A few centimeters between a tracked color patch and the rim is not enough; the gripper housing sits lower and farther out than the patch and will strike the edge. Do not use the interior-point Z as this gate. If clearance_z is below 0.30 m, lift first with a vertical EEF adjustment or adjust_height. Do not sweep the arm across the rim while raising. Only after current tracking shows at least 0.30 m, compute the planar residual and move XY. Keep consuming near_rim after every planar step; it remains the live edge height, not a one-shot pre-check. After XY is in band, descend; the lower bound becomes 0.20 m. Run the final Z stage below; do not skip it. After that descent, correct XY at the current height; do not lift back to the 30 cm planar-travel gate for that small correction.

Quadrilateral XY

Use this branch when floor_size=large. In the same atomic call as the held-object point and near_rim, bind:

• x1 and x2: two points on the intended container floor. Only their

robot-base X values are used. They mark the near and far feasible range along +X

• y1 and y2: two points on the intended container floor. Only their

robot-base Y values are used. They mark the left and right feasible range along +Y

The unused coordinate of each constraint point is ignored. Each constraint point must sit on the intended container floor, with Z clearly below near_rim and not on the room floor. After the bind, sort the live coordinates:

x_lo = min(x1_x, x2_x)

x_hi = max(x1_x, x2_x)

y_lo = min(y1_y, y2_y)

y_hi = max(y1_y, y2_y)

The four lines x = x_lo, x = x_hi, y = y_lo, and y = y_hi form an axis-aligned rectangle. Before using it, confirm from the current image that this entire rectangle is a subset of the intended container interior. Inset the four clicks from walls, partitions, and the approaching rim so the rectangle stays inside the intended bay. If the rectangle spans two drawers, covers a divider, or includes room floor or cabinet top, re-bind inset on the correct interior. If x_lo >= x_hi or y_lo >= y_hi, re-bind.

The XY target is x_lo < object_x < x_hi and y_lo < object_y < y_hi. The held object's current X and Y must lie strictly inside that rectangle. While this branch is active, do not also require the single-interior 2 cm XY band. A far floor point is one extent of the rectangle, not a reason to replace it with a click nearer the object.

Single-point XY

Use this branch when floor_size=small. In the same atomic call as the held-object point and near_rim, bind one stable point visibly inside the target opening. Prefer the container floor or an already-placed in-can object. For an ashcan, do not bind the far inner wall. Change viewpoint until the bottom disk or in-can soda is visible. For other small containers, bind the far inner wall only when a floor or in-can click is not obtainable. Do not use that Z as rim height.

Use object and container coordinates from the same current observation state. Compute the signed planar difference:

delta_x = container_x - object_x

delta_y = container_y - object_y

The classified surface picks the XY target band. Chassis or EEF motion may be used to enter the band; arrival is judged from current tracking, not from the commanded step size.

• interior_surface=far_wall: the target is delta_x of −3 to −5 cm

(-0.05 m to -0.03 m) and |delta_y| ≤ 2 cm. Do not release while delta_x is still positive. The held object must go a few centimeters past that far-wall point so it clears the near rim.

• interior_surface=floor: the target is |delta_x| ≤ 2 cm and

|delta_y| ≤ 2 cm. No far-wall bias.

Shared travel

While the classified XY band is not yet satisfied, keep clearance_z at least 0.30 m while moving XY. Do not close the remaining height until the classified XY band is satisfied. A visually centered footprint does not replace the classified XY band. After descent, correct remaining XY at the current height; do not lift back to the 30 cm planar-travel gate for that small correction.

Move only the holding arm using the largest collision-free portion of the current difference supported by visible clearance. A direct robot-base X/Y EEF call is appropriate when the arm can make the motion. A point-to-point or whole-body reach tool may also be used when its current tool contract, selected points, holding-arm choice, and clearance fit the state; do not ban a suitable tool merely because another recovery was previously preferred.

After every action, use measured post-action coordinates and the current image. Never update position by adding the requested displacement to an old value, reuse a stale residual, or issue an unmeasured sequence of corrections. If the object made safe measured XYZ progress, recompute from the new state and continue even when the controller reports timeout, non-convergence, ok=false, orientation error, or a residual. If the object did not make useful position progress, do not blindly repeat the same request. Reassess reachability, clearance, contact, and the current target geometry.

When the holding arm is at a workspace limit, use an evidence-supported chassis, body, or whole-body reach adjustment to restore reachability while preserving the grasp. If a vertical EEF lift cannot yet make the 0.30 m near_rim clearance, raise with adjust_height or another body adjustment that does not sweep the held object across the rim. After recovery, observe as needed, consume updated tracking, and recompute the complete remaining vector. Continue while distinct safe actions produce measured progress. Halt the current correction when a hard gate above applies; then keep seeking a different safe recovery while remaining in this Skill.

If planar adjust_*_eef_* shows the opening is still out of reach and the next step is to fall back to move_to_reach_point, first call set_arm_to_grasp_position with arm="both" and gripper="keep" so both arms return to the grasp-prep stow. The same stow is required before returning to the parent or any parent chassis reach after release, using gripper="open" on the unused placement arm. Do not start that chassis reach with an extended placement pose: the outstretched holding arm can knock the container or nearby objects. After both arms are stowed, re-observe, call spin_to_facing_point on the current target, then call move_to_reach_point only if the blue path on that post-facing image is still clear.

Align and release

Once the classified XY band is satisfied, descend. Do not release from a high hover.

Compute clearance_z = held_z - near_rim_z from current tracking. After XY is in band, the lower bound becomes 0.20 m. Adjust only Z until clearance_z is in 0.20–0.25 m. Never go below 0.20 m. If clearance_z is above 0.25 m, descend. If it is below 0.20 m, lift. Use a vertical EEF call or adjust_height. A pure-Z command can still change X/Y. After each Z change, consume current tracking and recompute the classified XY residual.

After descent, small XY corrections stay at the current height:

• If the classified XY band is no longer satisfied, or the held

object is clearly outside the opening, or the image shows the can sitting on the opening boundary or rim, immediately call track_object_distance on the current clickable head image_id and replace the full point set this Skill requires, including bottom interior and near_rim. Then correct XY at the current height; do not lift back to the 30 cm planar-travel gate for that small correction. A horizontal correction does not force another descent when clearance_z is already in 0.20–0.25 m. A drop that strikes the rim can tip the container and fail the parent task.

• If the image shows the held object aligned over the bottom

interior, not the rim, the classified XY band is still satisfied, and clearance_z is in 0.20–0.25 m, release.

Open only the holding gripper when the latest usable post-action evidence shows all of the following:

• the classified XY band is still satisfied
• the current image does not show the held object clearly outside the opening

or sitting on the opening boundary or rim

• the object's estimated downward footprint fits inside the visible opening

and is aligned over the bottom interior

• the gravity-aligned fall path enters the interior without striking the rim,

wall, handle, or existing contents

• the gripper has clearance to open without contacting the container
• current object/interior geometry does not contradict the visual judgment
• clearance_z is in 0.20–0.25 m and never below 0.20 m

Tracked point equality is not enough for release: points do not encode the object's footprint or the opening boundary. The classified XY band is still required and cannot be waived because the opening looks large. A grip point, centroid, visible narrow edge, or tracked interior point must not substitute for the complete downward footprint or sweep volume of an elongated object.

Release and verify

After open_gripper, inspect its current post-action image when usable; otherwise obtain a fresh head image. Allow settling evidence to guide timing. A mid-fall frame, a thin edge crossing the rim, or a centroid inside the opening is not enough to report this object complete. Wait for a settled observation of the complete object inside the container. If the arm occludes the result and current clearance supports it, make a small safe withdrawal with the former holding gripper open, then verify. That withdrawal is only for a clear view. It is not the required stow. Once the object is visibly settled inside, withdraw the arm to a safe pose if needed, then call set_arm_to_grasp_position with arm="both" and gripper="open" so both arms return to the grasp-prep pose. Do not exit this Skill until that set-arm has been called. Then report this object's placement complete. That exits this Skill only. Do not end the Codex session. Do not require reset_body as placement cleanup.

If the object is visibly outside the container after release, do not claim success. If it is still held, continue alignment and release from current evidence. If it is no longer held, first stow the unused placement arm with set_arm_to_grasp_position as required above, then return to the parent task. The parent chooses the next unused can. A can left outside after a miss is not that next can unless the parent last-step far-can rule says so. Do not immediately activate pick-up-object for this same object. Do not end the session from this Skill; the parent may end after its last-step 0.5 m check.

Treat a timeout or a response with missing completion details as an unknown post-action state. Do not blindly replay the command or release. First inspect the returned image or obtain a current observation, consume updated tracking, and compare measured coordinates with the previous state. Continue from that observed state when the grasp is safe and actual progress occurred. A retryable HTTP 504 is not by itself a transport failure. Keep inspecting the returned or newly captured state and continue this Skill. An unreachable interface or crashed simulator is the global session-end condition, not a placement exit.