把动作规则，写进阶段控制。

下载原文

# Layered memory index

Generated from memory leaf frontmatter.

## Global

- [Key: plate-normal grasp and rejected handover poses](global/insert_key_gpt_only_observations.md) — insert_key, GPT-only physicalrsi_v7 through v11
- [Latest v13 insert_key: reviewed native outcome and v14 hypotheses](global/insert_key_v13_outcome.md) — insert_key; read before physicalrsi_v14 pilot; seed0/layout0 evidence only
- [Latest v14 insert_key: reviewed native outcome and next hypotheses](global/insert_key_v14_outcome.md) — insert_key; reviewed v14 pilot; seed0/layout0 evidence only
- [Latest v15 insert_key: reviewed native outcome and next hypotheses](global/insert_key_v15_outcome.md) — insert_key; reviewed v15 pilot; seed0/layout0 evidence only
- [Key video review: paired presentation and recovery semantics](global/insert_key_video_review_v12.md) — insert_key; planning the next controller after physicalrsi_v12; hypotheses require implementation and native validation
- [Tubes: placement evidence, path ordering and tracking correction](global/insert_tubes_gpt_only_observations.md) — insert_tubes, GPT-only physicalrsi_v11 stock test_tube/00002
- [Latest v12: memory used; two placements, third orientation rejected](global/insert_tubes_memory_v12_outcome.md) — insert_tubes, GPT-only physicalrsi_v12, seed 0 / layout 0; read with earlier observations
- [Latest tube transport outcome and remaining grasp risk](global/insert_tubes_transport_grasp_v11.md) — insert_tubes; final v11 evidence for v12 memory iteration
- [Latest v13 insert_tubes: reviewed native outcome and v14 hypotheses](global/insert_tubes_v13_outcome.md) — insert_tubes; read before physicalrsi_v14 pilot; seed0/layout0 evidence only
- [Latest v14 insert_tubes: reviewed native outcome and next hypotheses](global/insert_tubes_v14_outcome.md) — insert_tubes; reviewed v14 pilot; seed0/layout0 evidence only
- [Latest v15 insert_tubes: reviewed native outcome and next hypotheses](global/insert_tubes_v15_outcome.md) — insert_tubes; reviewed v15 pilot; seed0/layout0 evidence only
- [Tube video review: reachable orientation staging and idle-arm clearance](global/insert_tubes_video_review_v12.md) — insert_tubes; planning the next controller after physicalrsi_v12; hypotheses require implementation and native validation
- [Evidence and recovery rules for coded insertion](global/insertion_evidence_rules.md) — GPT-only staged insert_key or insert_tubes, stock RoboDojo assets

## Suite

_(none)_

下载原文

---
id: insert_key_gpt_only_observations
scope: global
kind: failure
title: 'Key: plate-normal grasp and rejected handover poses'
applies_when: insert_key, GPT-only physicalrsi_v7 through v11
confidence: single-shot
evidence:
  cells:
  - v6_insert_key_s0_l0
  - v7_insert_key_s0_l0
  - v11_insert_key_s0_l0
  attempts: 3
reviewed: '2026-09-29'
provenance: Native results, stage traces and videos from the insertion campaign; human-reviewed
  synthesis.
---

## Verified observations, limited to this layout
The key plate is local X-Z and its normal is local Y. Closing on the circular bow edges along local X caused a drop after giver release in v6. A plate-normal receiving grasp remains the intended geometry; it has NOT yet been demonstrated in native simulation.
In v7 the required roll was rejected at the initial right TCP. In v11 moving the open right hand near [0.160,-0.208,0.926] succeeded, but BOTH equivalent jaw rolls were rejected by IK. This is negative evidence for the current fixed presentation/staging geometry, not proof that the task is mechanically impossible.
About 10.4 degrees of key settling occurred in the giver during receiver positioning. The controller now allows up to 20 degrees during receiving stages while preserving the original grasp reference, the 8 mm translation guard, measured jaw-normal checks, and independent lift verification after giver release.

## Current limitation / next hypothesis
The current template exposes no free handover pose parameter. Memory cannot authorize unavailable movement tools or force a rejected pose. A future controller may need a different giver presentation pose and a paired reachability check for both arms before contact. Do not claim that re-reading memory or switching a hole fixes this key kinematic limit.
For insertion only after VERIFIED transfer: key tip local -Z points down, directed tooth side local +X matches slot local +X before the final turn. Recompute geometry from live poses; never replay old XYZ coordinates as targets.

下载原文

---
id: insert_key_v13_outcome
scope: global
kind: failure
title: 'Latest v13 insert_key: reviewed native outcome and v14 hypotheses'
applies_when: insert_key; read before physicalrsi_v14 pilot; seed0/layout0 evidence
  only
confidence: single-shot
evidence:
  cells:
  - v13_insert_key_s0_l0
  attempts: 1
reviewed: '2026-09-29'
provenance: Reviewed native states, result, log, draft and video alongside PhysicalRSI
  reference. No automatic promotion of GPT prose.
---

## Verified interrupted outcome
v13 seed0/layout0 ended by explicit operator stop at native107,16 GPT turns. Native success was false. Giver pickup was independently verified at52. Presentation candidate tilt0/height0.89 passed nominal paired preflight and executed to84. Receiver translated to its staging position, then both equivalent rolls were rejected without motion at99. Preflight from the earlier joint seed did not guarantee reachability after the giver moved and physics settled.
At99 grasp error was5.019mm/11.370deg, accepted by receiver_orient/recover. Recovery independently lifted the key19.409mm; at107 error was5.375mm/11.926deg. The next present rejected the same still-held grasp under its old10deg limit without any new native action. This phase-dependent guard inconsistency is not evidence of fresh slip. Restart with the other arm then hit joint_path_rejected.
## Video comparison and next hypotheses
Reference key-selected at6–8s shows a receiver oriented across the exposed bow, with the giver releasing only after receiving. Our receiver remains near its original wrist orientation when roll planning fails. Images do not establish exact metric angles.
v14 implements consistent pre-handover present/recover tolerance, prefers tilted paired presentation candidates, and tries complete receiver staging poses from the current IK seed before translation-first fallback. These changes require fresh native validation. No successful handover, insertion or turning was demonstrated in v13; no successful-task recipe is published.

下载原文

---
id: insert_key_v14_outcome
scope: global
kind: failure
title: 'Latest v14 insert_key: reviewed native outcome and next hypotheses'
applies_when: insert_key; reviewed v14 pilot; seed0/layout0 evidence only
confidence: single-shot
evidence:
  cells:
  - v14_insert_key_s0_l0
  attempts: 1
reviewed: '2026-09-29'
provenance: Reviewed native states, result, log, draft and video alongside PhysicalRSI
  reference. No automatic promotion of GPT prose.
---

## Verified v14 interrupted outcome
Seed0/layout0,138 native actions,21 GPT calls,USD1.4684265. No native success; explicitly operator stopped. At52 pickup verified. Tilt-45deg/height0.89 presentation executed to87 with0.78mm/1.69deg grasp drift. Receiver roll was now executable: alternate direct staging executed34 steps to121. However measured TCP missed staging by about32mm and rotation was biased. Controller correctly blocked before approaching the bow. Recovery then re-presented the horizontal midpoint, shifting the key toward the lock; second staging missed by about64mm at137. Subsequent presentation candidate probes all failed; no actions at138. This is not a completed benchmark failure and no transfer/insertion was demonstrated.
## Next controller hypotheses
v15 adds up to3 checked open-hand pose corrections at the original staging goal, using measured position/rotation residuals. Per-increment and total offsets are bounded, lack of improvement blocks continuation, and the giver stays closed. Presentation retains its first successful XY instead of recursively moving the midpoint. These changes are untested natively. Preserve actual perpendicular jaw and independent lift checks. Do not claim a successful bow grasp from a completed planned path.

下载原文

---
id: insert_key_v15_outcome
scope: global
kind: failure
title: 'Latest v15 insert_key: reviewed native outcome and next hypotheses'
applies_when: insert_key; reviewed v15 pilot; seed0/layout0 evidence only
confidence: single-shot
evidence:
  cells:
  - v15_insert_key_s0_l0
  attempts: 1
reviewed: '2026-09-29'
provenance: Reviewed native states, result, log, draft and video alongside PhysicalRSI
  reference. No automatic promotion of GPT prose.
---

## Verified v15 interrupted outcome
Seed0/layout0;146 native steps,15 GPT calls,USD1.0211305. Native success false. Left pickup and tilted-45degree presentation succeeded. Right receiver staging reached an executed pose but its TCP missed by30.6mm on Y, total about31mm, with10.1degree rotation residual. One bounded compensation increased total staging error to38.38mm; the lack-of-improvement guard stopped it. Recovery and re-presentation retained the original XY as intended, but the next receiver correction again failed (37.37mm/9.01degrees). Operator stopped this repeated failure; it is an interrupted pilot, not a full benchmark failure.
The receiving right joint5 repeatedly settled near-1.55rad. Inspection of the actual x5/ARX.usd found joint1..5 limits approximately+/-10rad and joint6+/-3.14rad, matching the reported planner ranges. There is NO evidence of an extra joint5+/-90degree hard limit. Contact or self-collision remains an unverified hypothesis. Do not promote it to fact.
## Evidence for the next role experiment
The earliest staged v1 pilot on the same seed/layout explicitly restarted with arm=right, approached at59, grasped at74 and independently verified18.815mm key lift at80, with3.56mm/7.94degree relative change. This proves a right-hand initial table pickup in that older template, not a complete handover or the newer nearest-roll pickup. The PhysicalRSI reference uses the right giver and left receiver. A later right-arm restart AFTER the key was moved is not evidence against this original table pickup.
v16 prefers right giver/left receiver, while retaining explicit arm selection and checked restart. It tries the equivalent pickup half-turn only after an unexecuted plan rejection. All jaw-normal, actual staging, giver support and independent transfer lift checks remain. Test this changed direction rather than repeating the failed left-giver/right-receiver geometry; complete handover is still unvalidated.

下载原文

---
id: insert_key_video_review_v12
scope: global
kind: strategy
title: 'Key video review: paired presentation and recovery semantics'
applies_when: insert_key; planning the next controller after physicalrsi_v12; hypotheses
  require implementation and native validation
confidence: single-shot
evidence:
  cells:
  - v11_insert_key_s0_l0
  attempts: 1
reviewed: '2026-09-29'
provenance: Reference video frames, current source and native traces reviewed together.
  No new trial; hypotheses are not validated recipes.
---

## Observed in the reference video
Source: https://assets.kinetixai.cn/0928/insert_key-selected.mp4 and matching insert_key-wrist.mp4.
At about5–6s the right-side giver adjusts presentation. About6.2–8.4s the left-side receiver takes the bow; the giver opens. By9.2s the hands have separated while the key remains roughly horizontal; about9.4–10.8s the receiver changes to tip-down entry. These are playback times, not native steps. The video does not provide exact world angles, contact forces, or a metric handover offset.

## Current controller evidence
_present fixes the key long axis to world±X and its plate normal to world+Z, with midpoint translation and minimum height0.93m. Both receiver rolls failed at the resulting staging pose in v11. This only rejects that pose pair, not all plate-normal grasps. The demo starts with the key on the right; our v11 key is at x approximately-0.11m, so blindly swapping giver/receiver to copy the video is unjustified.
At native99 the giver's measured grasp error was4.576mm/10.3976deg. receiver_orient accepted it under8mm/20deg; recover rejected the IDENTICAL observation at the SAME native step because it returned to10deg, then cleared the verified grasp. This is a phase-dependent guard inconsistency, not evidence of additional physical slip. The current v12 controller still has that behavior.

## Untested improvements and limits
Evaluate paired giver/receiver presentation candidates, changing key position and plate tilt while retaining perpendicular jaw closure, instead of repeating rejected receiver rolls at a fixed key pose. Check receiving and later entry reachability before committing contact. After verified transfer, establish clearance from the open giver before large rotation, using checked paths and object-following evidence. Recovery should preserve the pending handover's meaning and independently reverify moderate settling before recalibration; do not broadly loosen insertion guards. These require controller changes; memory cannot call missing stages or establish a successful handover.

下载原文

---
id: insert_tubes_gpt_only_observations
scope: global
kind: strategy
title: 'Tubes: placement evidence, path ordering and tracking correction'
applies_when: insert_tubes, GPT-only physicalrsi_v11 stock test_tube/00002
confidence: single-shot
evidence:
  cells:
  - v6_insert_tubes_s0_l0
  - v9_insert_tubes_s0_l0
  - v11_insert_tubes_s0_l0
  attempts: 3
reviewed: '2026-09-29'
provenance: Native results, stage traces and videos from the insertion campaign; human-reviewed
  synthesis.
---

## Observed
Grasp the upper body below the cap and verify a lift. Tube0 and tube1 completed placement cycles in v6. Tube2 was verified upright, but crossed the occupied right-end tube while approaching a central hole; it tilted about 59 degrees and lost the grasp. Collision is the leading video/geometry explanation, without direct contact telemetry.
v9 raised above all other tubes, even off-route ones, and accumulated 3-4 mm execution bias during alignment. Repeating the same target did not remove that bias and consumed the native budget. v11 filters obstacles by the transport route and uses measured endpoint corrections. In the current v11 pilot, the first two tubes were released and stable by native step 308 without failed-stage recovery; the third is still under evaluation.
A placed tube may lean while satisfying fresh native conditions. Do not override fresh native containment/depth/upright evidence solely because root XY differs from the chosen hole center. Homing is deferred until all three tubes are released, then call the last tube's home stage again until both used arms are home.

## Untested next planning hypothesis
For this layout, try tube0 into central hole 2 first, tube1 into left outer hole 0, and tube2 into right outer hole 4. This may avoid carrying the final right-hand tube across an already occupied right-end hole and reduce required lifting/travel. This ordering is NOT a proven successful recipe. Inspect fresh locations and occupancy before choosing it; controller guards remain authoritative. Hole rows and coordinates are rack-relative, not fixed world poses.

## Record
For each tube retain object id, chosen hole, verified lift, alignment error, insertion/release evidence, and home status. Three inserted tubes without required arm home conditions still fail the full native task.

下载原文

---
id: insert_tubes_memory_v12_outcome
scope: global
kind: failure
title: 'Latest v12: memory used; two placements, third orientation rejected'
applies_when: insert_tubes, GPT-only physicalrsi_v12, seed 0 / layout 0; read with
  earlier observations
confidence: single-shot
evidence:
  cells:
  - v12_insert_tubes_s0_l0
  attempts: 1
reviewed: '2026-09-29'
provenance: Reviewed against result.json, states.json, run.log, GPT draft and native
  video; not automatic promotion of model prose.
---

## Outcome and scope
This supersedes any assumption that the center-first order is a validated complete recipe. The v12 pilot read MEMORY.md and two tube leaves before its first motion, chose tube0/hole2, tube1/hole0, tube2/hole4, and successfully updated its episode draft 11 times. It used 52 GPT turns and 397 native steps, then was deliberately operator-interrupted after repeated unreachable orientation plans. Raw termination_reason is infrastructure_error from the explicit operator stop, not an upstream API outage. It is an invalid/interrupted trial, not a complete benchmark failure or a success.

## Verified partial progress
Tube0 completed release, retreat and placement checks at native step 180 in central hole2. Its transport required independent grasp re-verification (22.32 mm lift), then a bounded recovery after the above-hole endpoint check failed. Tube1 completed in left hole0 at step332 without blocked-stage recovery. At final step397 both depth predicates were current and true; containment was last evaluated true at step388, and fresh axis measurements were about 0.38 and 7.29 degrees. Do not relabel the older containment evaluations as current-step evaluations.
Tube2 was grasped by the right hand and independently lifted 22.41 mm at step387, but remained about 87.8 degrees from upright. Its cap-up target was rejected for orientation at hole4, again after a recovery lift, and again after selecting hole9. All three failed orientation plans executed zero native actions. The other two tubes remained released. Neither arm met the required home condition; native eval_success remained false.

## Failed approaches and limits
Changing height with recover did not make this third-tube orientation target reachable. Changing hole4 to hole9 altered the yaw but also failed. Transport and handover could not bypass the required orient stage; restart was correctly blocked while the grasp remained verified. Do not use those rejected stage changes as a recovery recipe. The generic arm-change error mentions handover, but this tube phase does not expose a usable handover path.
The draft's suggestion to try another aperture is an untested hypothesis, not evidence that it would fix reachability. The intended final hole4 transport was never reached, so this trial does not establish whether the center-first order avoids the historical last-tube collision. Memory and transport re-verification code both changed in v12; this is not a memory-only ablation.

## Next controller hypothesis, not an executable recipe
Evaluate a supported held-tube translation toward a reachable orientation staging pose before rotating, or choose a different verified pickup transform. Check the candidate orientation before committing contact motion and retain lift verification and stage guards. No such change was tested in v12; memory cannot make an unavailable controller action callable. Budget native steps for all three placements plus both arms returning home. Only fresh native success can promote a full task recipe.

下载原文

---
id: insert_tubes_transport_grasp_v11
scope: global
kind: failure
title: Latest tube transport outcome and remaining grasp risk
applies_when: insert_tubes; final v11 evidence for v12 memory iteration
confidence: single-shot
evidence:
  cells:
  - v11_insert_tubes_s0_l0
  attempts: 1
---

## Final update to the v11 pilot
This supersedes the earlier note that tube2 was still under evaluation. The final native step was 443. Tube0 and tube1 were released and stable. Tube2 reached the high transport path, then about 9.76 mm translation and 5.42 degrees of relative grasp change tripped the 8 mm guard; it was no longer considered held. Its upright pose could not use the lying-on-table pickup template. The operator stopped unchanged retries. The task did not succeed and was an interrupted pilot, not a complete benchmark failure.

## v12 controller change to verify, not assume
Moderate settling (at most 20 mm / 20 degrees) during clear transport now requires a fresh independent lift before accepting a new grasp reference, as already done after rotation. Loss of following still blocks the stage. This has offline regression coverage but no native success evidence yet. Do not manually relax the insertion or grasp guards.
The center-first hole ordering in the prior tube strategy note remains an untested planning hypothesis. Reusing this memory does not itself establish improvement; inspect the actual choices, grasp evidence, native step use and whole-task verdict.

下载原文

---
id: insert_tubes_v13_outcome
scope: global
kind: failure
title: 'Latest v13 insert_tubes: reviewed native outcome and v14 hypotheses'
applies_when: insert_tubes; read before physicalrsi_v14 pilot; seed0/layout0 evidence
  only
confidence: single-shot
evidence:
  cells:
  - v13_insert_tubes_s0_l0
  attempts: 1
reviewed: '2026-09-29'
provenance: Reviewed native states, result, log, draft and video alongside PhysicalRSI
  reference. No automatic promotion of GPT prose.
---

## Verified interrupted outcome
v13 seed0/layout0 was explicitly operator-stopped at native493/500,55 GPT turns. Native success false; this is an interrupted pilot, not a full benchmark result. Tube0/hole2 and tube1/hole0 completed by180 and332. Tube2 was independently grasp-verified at387. A checked combined inward translation and cap-up rotation using staging fraction0.35 succeeded by418, where fixed-XY orientation had failed in v12. It reached above hole4 by447 and was inserted, released and stable by493.
At493 all nine native placement conditions were CURRENT and TRUE: depth, containment and cap-up for all three tubes. The both-arm home predicate was current and FALSE. Seven remaining steps could not cover guarded homing, so the whole task did not succeed. The tube2 insertion must not be described as complete benchmark success.
## Video comparison and next hypotheses
Reference at14–15s moves the tube toward the rack while orienting it, consistent with our successful inward staging. Reference left arm returns toward home after its last assigned tube at5.2–6.6s. Our final frame retains both hands above the rack. Frames cannot identify precise paths or establish contact force.
v14 uses the existing speed_scale1.25 only for tube approach/orient/transport free-space motions; independent lift verification and alignment/contact/insertion remain speed1.0. This is an untested step-saving hypothesis. Budget all three placements plus both homes. Earlier idle-arm homing is still a hypothesis, not an available stage bypass. No complete task recipe is published.

下载原文

---
id: insert_tubes_v14_outcome
scope: global
kind: failure
title: 'Latest v14 insert_tubes: reviewed native outcome and next hypotheses'
applies_when: insert_tubes; reviewed v14 pilot; seed0/layout0 evidence only
confidence: single-shot
evidence:
  cells:
  - v14_insert_tubes_s0_l0
  attempts: 1
reviewed: '2026-09-29'
provenance: Reviewed native states, result, log, draft and video alongside PhysicalRSI
  reference. No automatic promotion of GPT prose.
---

## Verified v14 partial progress
Seed0/layout0. Tube0/hole2, tube1/hole0, tube2/hole4, right/left/right arms. Free-space speed1.25 reduced third-tube grasp-verified step from387 to347. Inward staging fraction0.35 again made tube2 upright. All three were inserted, released and stable by447 versus493 in v13. Right exact-joint homing used40 more actions to487. Only13 actions remained for the left arm; native whole-task success remained false. Operator stopped future GPT calls after guarded homing was blocked. This is an interrupted pilot, not a whole-task success or full benchmark failure.
## Next controller hypothesis
v15 targets the full saved initial end pose through the existing checked Cartesian primitive at speed1.25 and verifies the current native per-arm home measurement. It does not require a specific elbow joint configuration. The generic reset primitive remains unchanged, as do all native placement and home thresholds. This aims to reduce conservative exact-joint sampling overhead; actual savings and success require fresh native validation. Maintain the working center-left-right hole order and inward last-tube staging. No complete success recipe yet.

下载原文

---
id: insert_tubes_v15_outcome
scope: global
kind: strategy
title: 'Latest v15 insert_tubes: reviewed native outcome and next hypotheses'
applies_when: insert_tubes; reviewed v15 pilot; seed0/layout0 evidence only
confidence: single-shot
evidence:
  cells:
  - v15_insert_tubes_s0_l0
  attempts: 1
reviewed: '2026-09-29'
provenance: Reviewed native states, result, log, draft and video alongside PhysicalRSI
  reference. No automatic promotion of GPT prose.
---

## One native successful pilot, v15
GPT-only openai/gpt-6-astra, memory enabled, max_turns80, stock ARX X5 assets, seed0/layout0. result.json is valid=true, success=true, termination_reason=native_success. All10 native conditions are current and true at490/500, including three depths, three containments, three upright checks, and both arms home.48 GPT turns,USD4.5325925. This is ONE successful seed/layout, not a measured benchmark success rate or a claim of robustness.
## Verified controller sequence
Use tube0/hole2 with right arm, tube1/hole0 with left, tube2/hole4 with right. Pick the upper body below the cap, independently verify lift, orient cap-up, clear other tubes before crossing, align the tube bottom to the chosen hole, insert in short checked segments, open in place, retreat and verify stable placement. Tube0 transport required bounded recovery and independent grasp re-verification; do not skip those checks. For tube2, fixed-XY orientation is unreachable in this grasp, but the checked inward staging fraction0.35 succeeds. All three releases/placement checks were completed by447.
Free-space approach/orient/transport use speed_scale1.25; independent lift, alignment/contact/insertion retain1.0. Empty-arm homing uses the existing checked Cartesian primitive at1.25 toward the FULL saved initial end pose, and requires fresh native per-arm home measurement. Right home used27 steps to474; left movement triggered native success after16 more to490, so the native simulator terminated before completing its entire requested initial-pose path. Generic reset still requires exact original joint angles, and no native judge or thresholds were changed.
## Scope and video evidence
The reference video supports cap-up placement and withdrawing an idle arm; it does not supply metric grasp offsets or prove complete final homing. Our final video and native states confirm three released tubes and the native home condition. Success followed controller changes plus accumulated memory; this is not a memory-only ablation. Recompute poses/apertures and retain guards in future layouts instead of blindly copying world coordinates.

下载原文

---
id: insert_tubes_video_review_v12
scope: global
kind: strategy
title: 'Tube video review: reachable orientation staging and idle-arm clearance'
applies_when: insert_tubes; planning the next controller after physicalrsi_v12; hypotheses
  require implementation and native validation
confidence: single-shot
evidence:
  cells:
  - v12_insert_tubes_s0_l0
  attempts: 1
reviewed: '2026-09-29'
provenance: Reference video frames, current source and native traces reviewed together.
  No new trial; hypotheses are not validated recipes.
---

## Observed in the reference video
Source: https://assets.kinetixai.cn/0928/insert_tubes-selected.mp4.
About2–3s and14–15s, a lifted tube moves toward the rack region while changing to cap-up. This supports considering position and orientation together; frames cannot distinguish one joint plan from several smaller commands. About5.2–6.6s the left hand leaves the rack and returns toward its initial region after its final assigned tube; the right hand handles the remaining two. The ending does not fully show right-arm homing.

## Current controller evidence
_orient preserves objectXY while rotating/lifting; lateral transport follows orient. In v12 tube2's measured grasp transform led to an orientation-constrained wrist target near x0.507m, rejected again after recovery and at x0.501m after changing aperture. The probe says position is reachable with the held orientation; do not describe this as a proven absolute x workspace limit. The required position/orientation combination failed. The current controller cannot translate a verified horizontal tube inward before orient through its public stage sequence.
At step332 the first two were placed, left hand still near the rack,168 native steps remained for the last tube and both homes. The center-first hole order did not reach the last tube's transport stage, so collision avoidance has not been validated.

## Untested improvements and limits
Plan a reachable free-space staging position for a held tube before rotation, or a checked translation/rotation path; account for the full horizontal tube, jaws, and placed tubes. Evaluate pickup transform together with downstream cap-up reachability. Preserve lift verification and slip guards. Assign arms and holes before execution; once an arm completes its final assigned tube, consider retreat/home instead of always postponing both homes until the end. Do not promise step savings without native measurement. Memory does not authorize stage skipping or declare successful task recipes.

下载原文

---
id: insertion_evidence_rules
scope: global
kind: strategy
title: Evidence and recovery rules for coded insertion
applies_when: GPT-only staged insert_key or insert_tubes, stock RoboDojo assets
confidence: single-shot
evidence:
  cells:
  - v6_insert_key_s0_l0
  - v6_insert_tubes_s0_l0
  attempts: 2
reviewed: '2026-09-29'
provenance: Native results, stage traces and videos from the insertion campaign; human-reviewed
  synthesis.
---

## Scope and evidence
All observations are seed 0 / layout 0 native pilots. No full-task success has been observed. Some runs were operator interrupted, so they are not complete benchmark failures. Only native eval_success establishes whole-task success.

## Apply
Read the current controller phase and fresh object poses. Track separate verified facts and hypotheses in STATE. A successful pose plan is not contact evidence. Gripper state is the previous control command, not measured finger contact. Object following during an independent lift is required to verify grasp/transfer.
Do not repeat a failed unchanged stage. Changing hole_index does not change a pickup pose, so it cannot repair the same unreachable pickup. A bounded recover is useful only while a grasp remains verified. Keep the original 300/500 native limits and reserve enough actions for required finishing/homing.
Published memory is read-only. Write a short episode draft after a new milestone or failure; include native step and observed evidence. Do not create a successful recipe for partial placements.

这是待核验草稿；结论需与原生结果对照。

Task insert_key: pickup -> handover -> insert -> turn. Controller physicalrsi_v16. Native155/300, eval_success false. Observed verified right blade pickup retained. Initial receiver staging at .89m failed46mm/9.97deg; recover raised key35mm, presentation .9182m passed paired preflight. Actual receiver staging failed again38.4mm/8.94deg; correction worsened commanded y residual to58mm. Right grasp residual .8mm/1.89deg. No handover yet. Hypothesis: another held-object recover raises staging into better actual reachability. Do not repeat unchanged receiver stage. Native requirements xy<7mm,upright<7deg,turn<30deg, depth threshold .025; no home final stage.

下轮可以读取；本轮启动时尚不存在，不属于本轮输入。

下载审核后的经验

---
id: insert_key_v16_outcome
scope: global
kind: failure
title: 'Latest v16 insert_key: reviewed native outcome and next hypotheses'
applies_when: insert_key; reviewed v16 pilot; seed0/layout0 evidence only
confidence: single-shot
evidence:
  cells:
  - v16_insert_key_s0_l0
  attempts: 1
reviewed: '2026-09-29'
provenance: Reviewed native states, result, log, draft and video alongside PhysicalRSI
  reference. No automatic promotion of GPT prose.
---

## Verified v16 interrupted outcome
Right initial table pickup with the current controller succeeded: approach44, close59, independently verified20.982mm lift at65. Presentation at-45degree plate tilt and0.89m height completed96. Left receiver staging then failed measured endpoint checks at138,155 and167 despite bounded correction. First corrected residual was45.99mm/9.97degrees; later residuals were38.41mm/8.94degrees and38.21mm/8.89degrees. Giver stayed verified throughout; no bow handover or insertion occurred. Native success false. Explicitly operator-stopped at167/300,18 GPT calls,USD1.304809. This is an interrupted pilot, not a complete benchmark failure.
Exchanging giver/receiver alone did not fix the staging problem. The left receiving joint5 settled near+1.54rad, similar in magnitude to the right's-1.55rad in v15. Actual USD hard limits match the planner's roughly+/-10rad on joint1..5; no extra+/-90degree hard stop was found. Contact or self-collision is not yet proven. Do not continue repeating the same presentation while assuming handedness alone solves it.
## Next geometry hypothesis implemented in v17
Vary horizontal key heading as well as plate tilt/height. Prefer bow heading35degrees toward the robot bases, with20,50 and0degree alternatives; these are controller candidate angles, not recovered from video. This places the open receiving wrist nearer its initial region while retaining approach along the blade and jaw closure perpendicular to the plate. Paired giver/staging/bow planning checks remain mandatory, as do measured convergence, giver support and transfer lift verification.
Record a presentation whose actual receiver-orientation stage fails and skip that candidate after recover while the grasp remains verified. Recompute from live poses, keep successful presentationXY fixed, and do not skip guards. Native success still needs validation; the v17 code is not a successful recipe yet.