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
- [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 tube transport outcome and remaining grasp risk](global/insert_tubes_transport_grasp_v11.md) — insert_tubes; final v11 evidence for v12 memory iteration
- [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_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_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: 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。

Insert tubes physicalrsi_v12 step397;103 native steps remain. Observed tubes0/1 released holes2/0 with native geometry passing. Tube2 right verified held sideways87.8deg. Recover selecting hole9 changed cap-up target to [.5012,-.1737,.966],quat[.2338,-.0188,.0048,.9721], but plan failed0steps, probe97%, orientation blocker. Hole4 also rejected previously; lift recovery alone insufficient; transport/handover skips and restart verified-held rejected. Hypothesis: another free rack aperture changes desired yaw sufficiently for planning. No overall success.

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_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.