32 current tasks: 28 retain provisionally, 3 retain with stated limits, and 1 review before final inclusion. Current catalog membership is synchronized; the additional shoe-removal recommendation has not been applied.
Dual shoes remain the main finalization concern: commanded gripper opening can trigger success before physical release finishes. Scanner retention includes the latest native success and ten static alignment checks. Block handover, nut assembly and trumpet cleaning retain the limits stated below.
All current measured verdicts and instructions remain intact. Single-seed results do not validate all configured instances or demonstrate RSI gains. No new model, GPU run or native replay was launched. Full review JSON · Number mapping · Report
| Current slot | Native task | Decision | Evidence and limitation |
|---|---|---|---|
| task01/01 | robocasa/LoadCondimentsInFridge | Retain provisionally | Retain the dynamic food identities and release/retreat clause. The terminal floor placements miss the storage goal itself. |
| task01/02 | robocasa/FilterMicrowavableItem | Retain provisionally | The sampled native labels identify the non-microwavable fruit and microwavable meat. Terminal evidence does not show the required bowl-and-meat containment. |
| task01/03 | robocasa/StoreDumplings | Retain provisionally | Keep side-by-side placement: the native receptacle helper requires direct contact, so stacking a dumpling on another can miss it. Containers remain outside the fridge in terminal evidence. |
| task01/04 | robocasa/DivideBuffetTrays | Retain provisionally | Group names must remain episode-specific. Food remains in the fridge; the failed result is not just a missing retreat. The initial camera does not show the whole workspace. |
| task01/05 | robocasa/MakeCheesecakeFilling | Retain provisionally | The lowered-head clause is necessary and already concise. The recorded speed dial remains off. No requirement to produce a visibly mixed substance is added. |
| task01/06 | robocasa/MultistepSteaming | Retain provisionally | Keep 'that same item': the native wording can call a sampled lemon wedge a vegetable. The source does not require turning on the stove or waiting for actual steaming. |
| task01/07 | robocasa/ScalePortioning | Retain provisionally | Retain release, waiting, and final retreat. The native weighing state is latched. The terminal food did not reach the complete weigh-and-plate goal. |
| task01/08 | robocasa/ScrubCuttingBoard | Retain provisionally | Several broad contact sweeps retain the native coverage condition; a single press does not. Remove the repeated explanation of broad coverage. Native success is geometric contact coverage, not simulated dirt removal. |
| task02/01 | libero/10_9 | Retain provisionally | Native BDDL byte hash matches the measured run. The yellow-and-white mug is distinguishable from the other mug. Both original clauses map directly to the goal; keep unchanged. |
| task02/02 | libero/10_3 | Retain provisionally | Native BDDL byte hash matches the measured run. The black bowl and bottom drawer are identifiable. Keep the original instruction unchanged. |
| task03/01 | robotwin/robotwin-aloha.handover_block | Retain with stated limits | Conditional retention only under the native final-state metric. The predicate has no history of left-to-right transfer; it cannot independently certify the requested handover. Do not weaken the instruction to turn this into plain placement. Gripper commands also do not prove completed physical release. |
| task03/02 | robotwin/robotwin-aloha.place_dual_shoes | Review before final inclusion | The native predicate DOES check both gripper-open flags, but those flags read commanded opening. In the measured journal, right-gripper opening is commanded at frame 1852 and success occurs at frame 1853 while the right finger joints remain about 0.0029 m, far below their approximately 0.045 m open pose. Physical release and stable hands-off placement are unproved. Recommend withholding this task from an unconditional final set under the unchanged predicate. It remains selected here because this request synchronizes the existing catalog rather than authorizing a new selection deletion. |
| task03/03 | robotwin/robotwin-aloha.put_bottles_dustbin | Retain provisionally | The initial three bottles and left bin match the wording. Terminal views show bottles in the bin. Remove the repeated destination sentence. The native center test is not full-mesh containment, hands-off stability, or a handover test. |
| task03/04 | robotwin/robotwin-aloha.scan_object | Retain provisionally | The instruction identifies the two held objects, the scanner optical window, the center of the box bottom, opposing faces, and an approximate 2–3 cm separation. Ten static physical-surface checks across all five scanner models pass the exact native predicate. Under the stated selection criterion, correct execution of this clear goal can pass; the verifier need not reject every incomplete pose. The live checker can trigger before the faces become parallel, so this native success does not independently verify precise face orientation or the literal surface gap. The task retains grasp stability, object-frame reasoning, inverse-kinematics constraints and two-arm alignment. One successful seed establishes a successful native attempt under this protocol; it does not estimate general success rate, prove instruction changes caused the improvement, or establish an RSI learning gain. It should no longer be described as unsolved in this latest run. |
| task04/01 | robodojo/make_toast | Retain provisionally | The four slices, two slots and rack are visible. Keep the count, one-per-slot, rack, lever and arm-return conditions. The task does not evaluate elapsed cooking or toast color. |
| task04/02 | robodojo/classify_objects_by_language | Retain provisionally | Keep the native reset-resolved category-to-basket mapping. A static generic override would erase the assignments. The catalog adds full insertion and arm return without fixing seed-zero categories. |
| task04/03 | robodojo/store_laptop_and_headphones | Retain provisionally | Keep earcup/headband orientation and fully closed, upright laptop seating. The latest headphone checks pass; laptop upright/stand checks fail. Wording can be shortened without deleting these goals. |
| task04/04 | robodojo/cover_blocks | Retain provisionally | Native reward checks six ordered stages. The wording already expresses the needed sequence and invariants; leave it unchanged. First and last views alone do not establish the history, so the sequential native verdict is essential. |
| task04/05 | robodojo/play_Xylophone | Retain provisionally | Keep the original left-to-right strike instruction. No native arm-return requirement exists. Strike detection is geometric, not a force or audio-quality metric. |
| task04/06 | robodojo/store_tools_in_toolbox | Retain provisionally | The shaped recesses identify destinations. Native helpers reject the latest slot alignments; hammer and pliers also miss full containment. Keep flat, matched outline, and below-rim seating while deleting repeated containment phrasing. |
| task04/07 | robodojo/insert_tubes | Retain provisionally | Pointed ends down and capped ends up match the saved-layout assets. Recorded insertion depths are about 54.75, 54.85 and 56.39 mm. Remove the duplicate seated-in-rack sentence while retaining deep insertion and arm return. |
| task04/08 | robodojo/deposit_coin | Retain provisionally | Keep the already revised instruction. It correctly requires full entry, without imposing bottom contact or resting. Native score is containment; it does not explicitly record the slot-crossing trajectory. The historical measured text differs from the current revised text. |
| task04/09 | robodojo/fasten_screws | Retain with stated limits | Retain with the existing observation caveat and explicit prior user selection. This is nut-on-bolt seating, not tightening torque or thread rotation. Red and blue assemblies pass; pink depth is only about 4.497 mm. Historical blue-object camera inconsistency is not repaired by instruction edits. Keep the current corrected instruction unchanged. |
| task04/10 | robodojo/play_stacking_toy | Retain provisionally | The saved-layout shape counts and peg heights support the mapping. Preserve all four group sizes, descending height assignment and precise seating; delete redundant restatement that pieces stay on the toy. The native success already requires every group, not merely partial credit. |
| task05/01 | behavior/assembling_gift_baskets | Retain provisionally | Native instructions and initial arrangement match the 16-object allocation. The run leaves the required distribution unfinished. Keep the native instruction rather than adding grasping recipes. |
| task05/02 | behavior/can_meat | Retain provisionally | The sausages are initially cooked; do not add a new cooking goal. The run does not complete the jar packing. Cabinet access and physical grasping are difficulty factors, not evidence of semantic mismatch. |
| task05/03 | behavior/moving_boxes_to_storage | Retain provisionally | The native disjunction allows either stacking order, matching the original instruction. The selected run uses the corrected house_double_floor_lower scene and achieves native success. Do not reuse the superseded scene-mismatch trial. |
| task05/04 | behavior/clean_up_your_desk | Retain provisionally | The object categories and fixtures match the original text. The existing run leaves organization unfinished; chair/access obstruction does not justify rewriting destinations. |
| task05/05 | behavior/setting_the_fire | Retain provisionally | The original wording captures placement, ignition and extinguishing the lighter. Burning wood outside the fireplace does not satisfy the goal. BDDL checks wood-in-fireplace explicitly; newspaper location is partly mediated by the wood-on-paper relation. |
| task05/06 | behavior/clean_a_trumpet | Retain with stated limits | Retain provisionally with visibility uncertainty. Latest complete evidence leaves 18 of 20 particles, so this is not a near-complete threshold-only failure. The older 19-of-20-cleaned interrupted run is not substituted. Keep 'all dust'; remove suggested turning and multi-view procedure. Visibility/reachability of the last particle across instances remains unproved. |
| task05/07 | behavior/slicing_vegetables | Retain provisionally | The original instruction explicitly asks to dice all five vegetables and close the refrigerator. The run fails during access before completing the transformations. Sampled camera frames do not expose the still-enclosed food; scene identities also come from BDDL and loaded-instance evidence. |
| task05/08 | behavior/freeze_pies | Retain provisionally | Keep the original packing and freezing goal. Frozen depends on temperature, not mere fridge containment. The terminal pies remain about 23 C and the manipulation goal is incomplete; this does not diagnose a cooling defect. Startup was completed in the selected run. |