[yuecideng]

Publication Update (2026-10-25): v36 And Main Synchronization

This section supersedes earlier current-head and qualification claims below.
The earlier notes are retained unchanged as historical provenance.

Published source head: 53b833c03c26c0891d808cc9c6129e2d409f1de9.
This is a fast-forward update from eb8871ae3, containing:

• 542c5caa9: the locally qualified v36 merge.
• 53b833c03: main 2f8aa96b84382d5525b9a19afbc1eb520bbc0f61 plus the
  resolved example-test conflict and full-pose TCP test-fixture compatibility.
  No production override was needed for that main merge.
• v36's bounded manipulation/recovery, directed-versus-undirected orientation,
  occurrence-local drawer contracts, and explicit UR10 native-Pick exploration
  remain in the GenSim adapter. The explicit UR10 ordinary-Pick experiment treats
  scene-mesh clearance as diagnostic, not certified. IK, command/velocity,
  tracking, effect and TaskSpec checks remain enabled. This publication check
  step makes no additional runtime or acceptance changes.

Current Native Qualification: Known Failures Remain

The frozen source head above ran 27 new full fresh public MiMo calls
(mimo-v2.6-pro, seed 0, one environment), with the same task inputs,
scene assets and declared configurations as the preceding v36 campaign.
There was no old model-response/bundle replay or manual retry in this campaign.
Historical scene revisions, per-case UR10 native-Pick opt-ins and the separately
authorized UR10 1176 physical reference were retained exactly as in the prior
campaign; this is not qualification on unmodified original scenes.

Robot / tasks	Program completion	Strict TaskSpec certificates
Franka / 1171-1190	15/20	7/20
UR10 / 1171-1177	7/7	2/7
Total	22/27	9/27

Franka tasks 1171, 1172, 1174, 1182 and 1186 did not reproduce program
completion. They remain unresolved and are explicitly retained for this PR
update; they are not reported as fixed or successful. The earlier latest-per-task
comparison was Franka 20/20 programs and 12/20 strict certificates, and UR10
7/7 programs and 2/7 strict certificates. That comparison includes the separately
authorized second attempt for Franka 1171; the new 27 calls are all first attempts.

For 1171/1172/1174, normalized graph/program/config declarations and compared
physical inputs agree with the prior runs. Main's effective full-pose IK
regularization changed, but this is a candidate variable, not an established
root cause. Fresh generated declarations differ for 1182/1186. No gate was
weakened to convert any of these outcomes into a pass.

All 27 saved videos decode completely. Human physical video acceptance is still
unreviewed; program completion is not a substitute for it or for strict
status=satisfied plus certificate_ready=true. Source, external-input and
scene fingerprints remain unchanged. Original prompts, failures, model receipts,
videos, and start/end/duration evidence remain in local ignored artifacts.

Latest Main / Conflict And CI Assessment

During regression, main advanced to
8cb2b53537e8c83b42a79c166e2777336ca0a2d0. It is not an ancestor of the
published source head. An isolated automatic merge of that exact main with this
head completed without conflicts; its resulting tree is
64224e6b10ae5bd7c9c6dd214947562ac68b77e6.
This prospective integration was validated without changing or committing the
already measured source branch.
After publication, GitHub generated merge commit
61fc7c2e5ff7f2ae5bead56cee502806ee8b1ee9 with parents main 8cb2b5353
and source 53b833c03. Its tree exactly matches the locally tested tree above.
GitHub now reports the PR mergeable (checks pending), not conflicting.

The latest main additions are expansion APIs and tutorial robot selection.
The reviewed defaults/import paths do not automatically enable a new GenSim
solver or acceptance path. No further main merge/rebase is performed here merely
to make the branch up to date; the observed branch rules do not require it.
Future main updates can still require renewed conflict/CI checks.

Validation performed for this publication:

• Prospective latest-main merge: 2759 passed, 25 deselected in the selected
  non-simulation/non-GPU/non-slow GenSim, TaskSpec, Task Program, expansion,
  robot catalog, package-data/examples and CI-contract selection. This is not
  the full repository suite or another native qualification.
• The identical prospective-merge selection also passed with isolated SciPy
  1.17.1: 2759 passed, 25 deselected, extending the previous fixture check
  to the v36 additions. This is not a claim that every CI dependency is identical.
• The previously failing pour-evidence selection: 15 passed with isolated
  SciPy 1.17.1 on the exact source head. The single-axis (6, 1) fixture fix
  is preserved. No production or dependency workaround was added.
• Source head: Black 26.10.1, 1748 files unchanged; API 3436/3436.
• Prospective merge: Black 1756 files unchanged; API 3474/3474.
• Context-map validation passed; 42 project-context tests passed.
• Additional independent checks: CI contracts 108 passed on each tree;
  motion/default-configuration CPU probes 15 passed across both trees.
• Whitespace and publication-blob checks passed; no experiment outputs,
  credentials or new oversized publication blobs are included.
• The prior exact-source check (1504 passed, 44 skipped, 2 deselected) remains
  separate evidence from the new prospective-merge CPU selection.
• Three independent bounded reviewers completed agentic, engine and verification
  passes with architecture/API/robotics/security lenses. No newly confirmed
  actionable code defect remained. Known native non-reproductions remain an
  explicit qualification limit, not a zero-risk claim.
• Existing behavioral and boundary coverage was reused; no new redundant tests
  were added for publication. Mapped context remains accurate, so this step
  requires no source/context edits.
• actionlint was unavailable; a full wheel build, full Sphinx build and full
  hardware matrix were not run locally. Packaging staging/contract tests were
  included in the selected CPU checks.

Remote CI must run against the new head and GitHub's current integration
commit. Old-head green checks do not certify this update. The push triggers new
checks; their success is not asserted here. This publication performs no force
push, PR merge, remote approval, or review event.

---

CI Compatibility Follow-Up (2026-10-24)

Current published head: eb8871ae300db415a0ed7ef9df31e6978108c12f.

The previous head's fast lane failed 15 pour-evidence tests at their shared
fixture, before observer execution: SciPy 1.17.1 requires an explicit
single-axis batch dimension, while the local SciPy 1.15.3 accepted shape (6,).
Commit eb8871ae3 changes only that test input to shape (6, 1). Production code,
acceptance thresholds, dependency declarations and all existing test cases
remain unchanged.

Validation:

• Reproduced the original ValueError with an isolated SciPy 1.17.1 installation.
• SciPy 1.17.1: 649 passed, 2 skipped in the Task Engine package, including
  all 15 previously failing pour-evidence cases.
• SciPy 1.15.3: 15 passed in the pour-evidence module.
• Black 26.3.1: 1726 files unchanged; API docs 3399/3399; diff checks passed.
• Existing behavioral coverage was reused; no duplicate tests were added.
• New CI is triggered by the fast-forward push. Its result is not yet claimed.

The earlier publication and native-evidence notes below remain historical;
this follow-up performs no new physics execution or task-generation requests.

Previous Publication Update (2026-10-24)

Published head: f1b750199d6df102f994057bc05a8544af4808ba. This is a fast-forward update from
3e002eea4, including the three previously saved convergence/acceptance commits,
the bilingual graph viewer (26d59250a), and the placement-evidence fixes
(f1b750199). The validation and head references in this section supersede the
older publication notes retained below.

Changes

• Move application-specific hand targets and supported-end grasp selection into
  GenSim while retaining shared hooks and default behavior. Run region selection
  after collision qualification and before ranking/top-K.
• Remove unused staged-yaw and settling-XY functionality and the legacy
  embodichain.task_spec forwarding namespace. Use
  embodichain.gen_sim.task_spec; retain existing world_yaw_free behavior.
  Changed execution declarations require regenerated bundles, not reused identities.
• Add measured pouring/support/coordinated-placement evidence and same-episode
  legacy/current observer audits, keeping execution and strict acceptance separate.
• Integrate the generic bilingual semantic graph viewer in the normal publication
  paths. Package and embed styles, JavaScript, fonts and licenses for offline use.
  Model-generated text is digest-bound, cached display metadata only; failures
  retain source text. Pure render/write/visualize APIs stay offline.
• Preserve explicit missing-preview placeholders for imported/copied geometry
  without source images; do not substitute unrelated images as evidence.
• Fix two independently reproduced P2 defects: an E5 contact timeline cannot
  replace an E1 supported-region footprint/contact-gap/release-window contract;
  a cavity-unknown environment row no longer invalidates other final or milestone
  rows. Matchers recompute row availability and reject forged masks.

Validation And Review

• 1145 passed, 2 skipped in the focused Task Engine, package-data,
  configured-integration, Atomic Action and settling selection. GPU tests remain
  skipped; this is not the full repository suite.
• 42 project-context tests passed; context validation passed.
• Black 26.3.1: 1726 files unchanged. API docs: 3399/3399 exports
  documented. Commit diff whitespace checks passed.
• Browser checks cover desktop/mobile, both languages, real success/failed/
  unavailable historical records, graph interaction, video, JSON/SVG export,
  injection resistance and no external requests. Package resources were exercised
  through real setuptools staging; no complete wheel build is claimed here.
• New P2 regression selection fails against the old implementation
  (3 failed, 2 passed) and passes after the repair (5 passed).
  Mixed-row tests retain known success/failure and preserve unknown as unavailable.
• Three independent bounded publication-review passes covered architecture/API,
  agentic/acceptance and security/resource boundaries. Two independent follow-up
  passes verified the P2 fixes. Test necessity was reviewed; no unresolved
  actionable finding remains in this publication increment.

Priority	Location	Defect	Impact	Action
N/A	N/A	No unresolved actionable findings	N/A	Both reproduced P2 issues fixed and regression-tested

Historical Evidence Boundary

The nine saved native episodes for tasks 1171, 1172 and 1181-1187 were re-evaluated
offline against the pre-fix and repaired evaluators using the same saved
measurements and runtime receipts. Every final-goal status, milestone status and
certificate flag was unchanged; all recorded input hashes were unchanged.
This preserves the existing seven strict certificates, task1184 unavailable and
task1186 failed results. It does not turn nine completed programs into nine
strict successes. The older native qualification table below describes a
different, earlier frozen qualification epoch and is retained as historical
provenance, not the current observer result.

No new physics, GPU task execution, motion commands or task-generation model
requests were used for the P2 repair verification. The graph translation smoke
test was a separate display-only model request. Historical specs, videos,
receipts and failures were not rewritten. Local evidence artifacts are retained
outside the Git publication.

Remote CI has been triggered for the new head and must complete independently;
old-head CI success does not certify this head. No force push, main merge, PR
merge, or remote approval/review event was performed.

Historical Publication Notes (Through 3e002eea4)

Description

Integrate GenSim with the regular Task Program and Atomic Skill runtime using
invocation-scoped policies, measured manipulation evidence, bounded recovery,
and explicit motion contracts. Keep shared Lab defaults separate from GenSim's
fixed-grid, sequential-seed motion adapter.

This update includes main 658377f65 and the qualified v36 task fixes at
b2a8b115f, merged as be90d1833. Publication head 3e002eea4 adds only focused
test cleanup; production code is unchanged from the nine-task native replay.

• Preserve execution-occurrence identity independently of cached analysis IDs.
• Align held-object commands, mixed held/free Place rows, payload-aware
  transport, and bounded grasp/placement/articulation recovery.
• Preserve ordinary motion, collision, joint-limit and velocity qualification;
  report experimental completion separately from strict TaskSpec acceptance.
• Calibrate qualified light pouring grips without changing unrelated hands or
  shared robot defaults; preserve Press withdrawal targets during post-hold.
• Support Franka, UR10 and Tianji Marvin / Robotiq bundle preparation and
  dependency preflight without replacing an existing bundle's embodiment.
• Require measured container landing geometry rather than object-name fallbacks.
• Restore offline semantic-graph HTML and committed-video reporting while
  retaining the shared full-episode audience camera.
• Preserve positional compatibility for ContainerAffordanceBinding by making
  the added release_clearance field keyword-only.

Refs #531

Type of change

• Bug fix
• Enhancement / new feature
• Breaking change
• Documentation update

Compatibility and dependencies

• Python imports under embodichain.gen_sim.simready_pipeline.* move to
  embodichain.gen_sim.asset_engine.*, without an old-namespace shim. The
  embodichain simready CLI name and flags remain available.
• ArticulatedGenerationClient moves from
  scene_engine.clients.articulated_generation to
  asset_engine.clients.articulated_generation under embodichain.gen_sim.
• The GenSim optional dependency group includes shapely, yourdfpy>=0.0.60
  and python-fcl>=0.7.0.11. This publication did not modify the host environment.
• Install the current editable project or wheel so bundled tasks, robots and
  the main branch's benchmark extension resolve correctly. Packaging retains
  Task Engine YAML/HTML and Asset Engine JSON resources.
• The packaged Task Engine CLI defaults to experimental program completion.
  Auto-selected Contact/Composition profiles may enable experimental release
  observer revision 2. Generated relative-placement post-policies can use the
  existing experimental 0.10 m position tolerance. These are not shared Lab
  defaults or substitutes for strict TaskSpec certification.
• Frozen bundles retain their declared robot, identity, physical inputs and
  motion contracts. Changed declarations require regeneration and validation;
  old receipts do not certify changed inputs.

Validation

• Current publication tree: 4046 passed, 50 skipped, 21 deselected in the
  focused software selection spanning GenSim, Task Program simulation and
  language/runtime integration, Atomic Skills, affected motion/object modules,
  supported-end grasps, package-data/installed-wheel checks, CI contracts and
  project-context checks. This is not the full repository/GPU suite.
• Black 26.3.1: all 1715 files unchanged after cleanup.
• Public API documentation: 3397/3397 exports documented.
• Context-map validation and git diff --check passed. Existing context
  accurately describes the merged behavior; the final test-only cleanup needs
  no production-context change.
• Three independent review passes found no unresolved publication-blocking
  issue. Test necessity was reviewed across all 80 changed test files in
  the full proposed PR, including extensions to existing tests and fixtures.
• Four test-quality issues were corrected: remove an identical Twist parameter,
  combine duplicate AxisAlign and articulation-descriptor cases while retaining
  their distinct assertions, and use independent fixed scene-category inputs
  rather than deriving the oracle from the production set.
• The resulting test delta is 12,572 added / 82 deleted lines. This replaces
  the older description's no-longer-current claim of fewer than 10,000 added
  test lines; no safety coverage was deleted to meet an artificial line count.
• Workflow contract tests passed. actionlint was unavailable locally; no
  full Sphinx build or additional full GPU suite was run in this publication step.

Frozen native regression

The merged production commit be90d1833 reran the same nine previously
qualified bundles with Franka, seed 0 and one environment, serially through
the existing resource supervision. There were zero new model requests.
All 183/183 compiled calls completed, every segment's validation mask was
accepted, and all source/bundle/external-input hashes remained unchanged.
Strict outcomes were read from the separate TaskSpec reports and match the
pre-merge reference outcomes.

Task	Completed / compiled calls	Separate strict TaskSpec	Certificate ready
1171	15 / 15	satisfied	yes
1172	12 / 12	failed	no
1181	15 / 15	satisfied	yes
1182	14 / 14	unavailable	no
1183	43 / 43	satisfied	yes
1184	43 / 43	failed	no
1185	12 / 12	unavailable	no
1186	15 / 15	failed	no
1187	14 / 14	unavailable	no

This establishes 9/9 complete experimental programs and 3/9 strict
certificates, not nine strictly successful tasks. Task1172 uses the previously
qualified 195 mm drawer asset; the same-named original baseline asset is not an
equivalent input. These runs do not establish fresh-provider generation,
cross-seed/scene generalization, or complete UR10/Tianji native qualification.

The final 3e002eea4 changes only four test files. All 1144 recorded production
source files still match the native replay epoch, so no further native replay
was required for that cleanup. Local receipts and videos are retained but are
not uploaded as repository outputs.

Publication hygiene

• Fast-forward update of this PR's existing source branch from e130ee3b5 to
  3e002eea4; no force push or history rewrite.
• Both main 658377f65 and v36 b2a8b115f are ancestors of the published head.
• Existing branches and rollback backups remain intact. No outputs/,
  credentials or oversized new blobs were introduced by this update.
• Existing labels remain applicable. This update does not merge the PR.
• Remote CI must run against the new head; success on the old head is not
  evidence for the new publication.

Screenshots

No new screenshots are attached. This is a runtime/integration and test update;
native recordings and offline evidence remain local artifacts.

Checklist

• I have run black . with Black 26.3.1.
• I reviewed affected documentation and agent context and explained why no
  additional context edit was needed for the final test-only cleanup.
• Public API documentation coverage passes.
• I verified behavioral coverage and reviewed added/extended test necessity.
• Dependency and installed-package impacts were reviewed.
• Remote CI for the updated publication head has completed successfully.

[greptile-apps]

Held continuations cannot qualify

For an E5 hold followed by an E5 on-placement using step_result, the transport call emits only move, hold, release, and retreat. This observer requires a lift event, so phase("lift") raises even when transport and release succeed. Both acquisition and placement evidence become unavailable, preventing strict TaskSpec acceptance.

Read grip evidence from the continuation's own move window, or bind it to the preceding acquisition, instead of requiring another lift.

Prompt To Fix With AI

This is a comment left during a code review.
Path: embodichain/gen_sim/task_engine/_task_program/coordinated_placement_evidence.py
Line: 229

Comment:
**Held continuations cannot qualify**

For an E5 hold followed by an E5 on-placement using `step_result`, the transport call emits only `move`, `hold`, `release`, and `retreat`. This observer requires a `lift` event, so `phase("lift")` raises even when transport and release succeed. Both acquisition and placement evidence become unavailable, preventing strict TaskSpec acceptance.

Read grip evidence from the continuation's own move window, or bind it to the preceding acquisition, instead of requiring another lift.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[greptile-apps]

Final placements stay unverified

When an E5 on-placement is the last group, it cannot pass its ordered milestones—even with valid grip, release, and support evidence. task_spec_from_candidate requires s1.acquired before s1, but evaluate_runtime_evidence first exposes both observations in the same group-end snapshot. _evaluate_milestones checks s1 starting at the next snapshot, which does not exist. Strict acceptance therefore remains unavailable.

Emit separate snapshots for the measured acquisition and placement boundaries.

Prompt To Fix With AI

This is a comment left during a code review.
Path: embodichain/gen_sim/task_engine/_task_program/task_spec_runtime.py
Line: 2509-2514

Comment:
**Final placements stay unverified**

When an E5 on-placement is the last group, it cannot pass its ordered milestones—even with valid grip, release, and support evidence. `task_spec_from_candidate` requires `s1.acquired` before `s1`, but `evaluate_runtime_evidence` first exposes both observations in the same group-end snapshot. `_evaluate_milestones` checks `s1` starting at the next snapshot, which does not exist. Strict acceptance therefore remains unavailable.

Emit separate snapshots for the measured acquisition and placement boundaries.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

[skywhite1024]

Main 同步与抽样验证

已推送 merge 提交 67b1b2b2e24ceb0ed1f4689040fc73f22d5f1eef，合入 main f3a399ec0cbca5aaf67c3783a81255a401a1103c。

• 唯一人工冲突位于 docs/architecture/curated.json，保留双方架构节点；未通过覆盖一侧解决冲突。
• 提交源码树为 30623dbeb53e0f9e567fc2d0509c737392f05bdb，与抽样验证的干净合并源码树一致。
• 原工作区三份未提交 DLS 候选修改不在本次提交中，未调整任务输入、种子或验收门槛。

验证：

• Black 26.10.1：1842 个文件无需修改。
• 定向软件测试：412 项通过、4 项跳过；另显式运行 CPU/Gloo DDP 测试，5 项通过。复用现有覆盖，没有为本次架构节点合并新增重复测试；未运行全量测试。
• API 文档覆盖：3572/3572；上下文检查与 diff 检查通过。
• 提交固定版本的架构生成通过：121 个节点、244 条关系。
• 构建完整 wheel 并隔离安装；未更新共享 conda 依赖。合并后刷新项目本地元数据，当前工作区 Franka/UR10 资产预检通过。

最新合并源码的完整 MiMo fresh 抽样（当前 baseline 的 new3 场景、历史实际提示词、seed=0、单环境）：

任务	程序	严格 TaskSpec	完整 CLI 时间
dual_franka / task1173	12/12	satisfied，certificate_ready=true	205.861 s
dual_ur10 / task1177	43/43	satisfied，certificate_ready=true	341.396 s

UR10 task1177 沿用历史 native_pick_exploration=True，不是默认模式认证。两题均重新调用 MiMo，无旧响应、图或轨迹重放，无 physical reference。结果仅覆盖抽样，不代表全部历史任务重新通过；视频完整人工验收尚未完成。

未运行完整 Sphinx 构建；本地没有 actionlint，相关 workflow 测试已包含在定向检查中。GitHub 已重新触发 CI，发布时尚未全部完成，不能将本地检查写作新 CI 全绿。

CI 后续状态：lint、Documentation tests 和 offline tools 已通过。test 在下载 test-plan 产物时失败，直接错误为 Failed to ListArtifacts: ... ECONNRESET，尚未进入 pytest 执行；这是本次失败日志确认的最早失败点，不能据此判定 task 或源码回归。已尝试单独重跑该 job，但 GitHub 返回 cannot be rerun；当前 docs build 仍在运行，重跑尚未触发，需待当前 workflow 结束再操作。未为该网络问题修改源码或放宽 CI 门槛。