Include callable skill and live-GUI prerequisites, epoch sensor reads, packed snapshots, sampled bit-exact kinematics, and verified-tick rendering without changing the physical/control schedule. Record bounded analytic S1 GUI/headless qualification, original failures and CPU evidence (1557 passed, 6 skipped). Preserve accepted_skill=false and sub-real-time limitations. Exclude native authorities, caches, large media and local delivery copies; retain sealed artifact digest.
15 KiB
Go2 FL仿真功能接口:0.4.1实时GUI扩展(非严格accepted_skill)
0.4.1:实时窗口与headless明确分开
公开Python增加严格布尔gui=False参数;CLI增加--gui。默认仍无窗口,控制/物理范围与0.4.0完全相同。示例:start(NEW_OUTPUT, NEW_REGISTRATION, sequence='S1', controller='analytic', gui=True)。gui进入精确spec、approval及认证FD,worker仅RUN使用AppLauncher(visualizer='kit', headless=False);PREP始终headless/零环境/零物理。GUI依赖在PREP/RUN预导入并绑定,仅演示配置增加KitVisualizerCfg,不改物理参数、原render_interval20或控制步长。
独立新包的bash demo_gui.sh是一条实时GUI启动命令:真实Kit窗口,默认解析S1。bash demo.sh仍真实仿真但没有窗口;bash verify_functional.sh为原完整headless功能验证。demo/measured_state_animation.gif若查看0.4.0中的该文件,则只是历史离线重建,不是GUI也不是Kit摄像机录像;0.4.1真实截图/短录像路径见该包README。
界面显示真实模式、命令、6相位及进度、周期与tick、接触Fn/gap/足端位置。相机/灯光只修改新视觉prim;蓝色3kg载荷是原资产刚体,不额外创建假载荷。原500Hz安全/记录先于25Hz渲染更新,KitVisualizer在app.update期间禁用player物理,保持窗口响应而不额外步进。
完整序列或显式安全停止后,query().state在查看期为viewing_final;这不是最终监督通过。保留原场景/最终画面(最多10分钟),无额外物理步;关闭主窗口或Close按钮后才完成preclose身份核对、实际wait和清理。若先到600秒查看上限,则有界退出并报告view_timeout,不把它标成用户关窗;原总截止/存储等guard优先,若它们更早触发仍按原规则失败退出,不保证自动查看超时总会先到。提前关窗会中止并保存部分轨迹、非零退出,原因gui_window_closed_before_completion(不等待下一周期、不冒充protocol_complete)。显式公开request_stop()仍是原完整reload边界的安全停止,不等于关窗或完整矩阵通过。
GUI receipt单独记录真实render更新、完成标志、关闭原因、最终查看carry/counter不变;截图/桌面可见性仍须外部实测,receipt本身不证明窗口真的可见。启动早期AppLauncher尚未返回时的关窗尚未实测;任何缺失worker/完整证据均不能通过监督。0.4.0全部原包/证据保留;新版本只对本轮实际列出的结果重新验证,不自动继承为全部GUI命令/u256资格。
原有功能API与约束(继续保留)
This is the existing aided, floating-root Go2 FL single-leg simulation with the original IK, contact logic, residual bound, fixtures and functional/safety gates. Analytic is zero residual; u256 is the unchanged deterministic frozen actor plus that analytic reference. No training. CPU qualification is not physical validation.
Public entrypoints (no Kit in caller)
From the repository or a same-layout independent copy:
from scripts.go2_fl_skill import start, SkillRun
run = start('/absolute/NEW/run', '/absolute/current-task/registration.json',
command='x_bias-16mm', cycles=6, controller='analytic')
status = run.query() # detached JSON snapshot, no tensor aliases
request_id = run.queue('y_bias-24mm') # one outstanding request; may raise FileExistsError
stop_id = run.request_stop() # request, not immediate termination
exit_code = run.wait() # waits for supervisor, which waits/cleans Kit
# Existing process: SkillRun('/absolute/NEW/run').query()/queue()/request_stop()
cycles is an exact integer 1..20 (not bool). IDs are validated_nominal (+25,+20 mm) and {x_bias,baseline_direction,y_bias}-{16,24}mm. Directions are normalized world-XY ratios (25,10), (25,20), (15,20). No arbitrary vectors, durations, controllers, resets or randomization. Controllers are analytic or u256. Physical u256 execution requires complete current-source nominal20/S1/S2/S3 analytic successes supplied using --analytics-evidence. Prerequisites must have full requested counts and exact fixed pair-command history; an early safe stop or changed schedule never qualifies. Both controllers accept the same finite command/cycle interface. Historical 0.4.0 physically qualified u256 only on S1, not every combination or 20 cycles; 0.4.1 has not requalified u256/headless or u256/GUI. Its current-source analytic prerequisites remain mandatory.
sequence= supplies the entire initial command/cycles/queue; do not combine it with custom command/cycles/queue overrides. Frozen schedules:
nominal20: nominal, 20 continuous cycles.S1: x_bias16 → baseline_direction24 → y_bias16, six cycles.S2: y_bias24 → x_bias24 → baseline_direction16, six cycles.S3: baseline_direction24 → y_bias24 → x_bias16, six cycles.
A is the initial commanded ground target, not measured landing; root target remains the initial commanded port. Each command owns an A→B(c)→A pair. Queued commands append FIFO (ten pending maximum) and apply only after an admitted return-to-A endpoint (completed cycles 2,4,...). Initial cycle zero is set by launch arguments; requests arriving during preparation queue for the next A endpoint, not an extra initial reference refresh. A pair's B never changes mid-pair, targets never accumulate, and history remains available for every recorded sample. Queueing never extends the requested cycle limit; stopping or reaching that limit can leave queued commands unexecuted. Stop is honored at any complete, audited, reloaded cycle endpoint, before seam admission/reference refresh; immediate safety failure always wins. Early completion is requested_stop, not protocol_complete for the requested maximum.
Invalid commands are rejected before any queue/control mutation. Runtime also validates exact mailbox fields. One atomic outstanding mailbox request, max 1024 input bytes, 64 receipts/run; malformed/overflow input cannot become control. Client rejection does not create mailbox files. Runtime rejection receipts may update audit metadata, never queue/control state. The mailbox is a bounded same-user local seam, not a hostile-user security sandbox.
query() reports state, authoritative global/cycle/local ticks, actual phase/progress, latched contact/normal force/gap, measured foot, commanded target/fixed anchors, pending/active IDs, stop request and failure. Physical status is copied every 500 physics ticks (one simulated second), at initialization, at endpoints and on failure; its tick identifies freshness. receipt.json records the latest consumed request immediately; last_receipt in status is sampled. Do not mistake starting for initialized physical state, or stopped for supervisor success. Final success needs supervisor_success=true and the successful waited campaign result; early stop retains requested/completed counts separately. If the API-owned supervisor exits before mailbox creation, its handle reports failure.
CPU validation only
/usr/bin/python3.12 scripts/go2_fl_skill.py validate
CUDA_VISIBLE_DEVICES='' /home/cen/miniconda3/envs/linker_sim_isaac/bin/python \
scripts/go2_fl_skill.py validate --verify-frozen
The second command calls unchanged replay.verify_bundle, including all 134 immutable snapshots/live dependencies and the exact u256 hash; only the preexisting AGENT.md exception remains. Neither command queries GPU, creates authority, launches Kit or advances physics.
Parent's real run commands (fresh authority generated automatically)
For a future explicit local invocation, create a fresh finite session first (CPU only, no GPU/Kit):
/usr/bin/python3.12 scripts/go2_fl_skill.py authorize --registration /absolute/NEW-session/registration.json
Python equivalent: scripts.skill_launch_v4.new_registration(Path('/absolute/NEW-session/registration.json')). This is explicit caller authorization, not automatic reuse/extension of the development task. The new session is capped at12 hours and retains the same launch/call/storage limits. A new explicit invocation uses its own fresh registration and deadline. It does not revive the expired 0.4.0 task or the completed 0.4.1 development task.
Use system Python3.12 for supervision. Each run creates NEW <output>-authority/approval.json, binding exact current source hashes, spec, fixed limits, current boot, current allowed desktop identities, admission window, registration hash and deadline. It then runs a new zero-physics config-preparation Kit and one physical Kit, serially. No old approvals, binding preparations, failure directories, resumes or retries are reused. Preparation is not physical qualification.
Choose never-used output names; these are examples, not an instruction to reuse an existing root. Shared registration is the parent's current task registration, not an old physical grant. Keep it identical across repository and independent-copy runs so they share skill-v4.lock/skill-v4-ledger.json in its parent directory.
set -eu
REPO=/path/to/current/repository-or-independent-copy
REG=/absolute/NEW-session/registration.json
P=/home/cen/sf-runs/skill-v4-NEW-PREFIX
CLI="$REPO/scripts/go2_fl_skill.py"
/usr/bin/python3.12 "$CLI" run --registration "$REG" --output "$P-nominal20" --sequence nominal20
/usr/bin/python3.12 "$CLI" run --registration "$REG" --output "$P-S1" --sequence S1
/usr/bin/python3.12 "$CLI" run --registration "$REG" --output "$P-S2" --sequence S2
/usr/bin/python3.12 "$CLI" run --registration "$REG" --output "$P-S3" --sequence S3
# Parent reviews all analytics before executing this gated run:
/usr/bin/python3.12 "$CLI" run --registration "$REG" --output "$P-u256-S1" --sequence S1 --controller u256 \
--analytics-evidence "$P-nominal20" --analytics-evidence "$P-S1" \
--analytics-evidence "$P-S2" --analytics-evidence "$P-S3"
# Public transport safety-stop probe: request after first 5 ticks, reserve 3 cycles; expect endpoint1.
/usr/bin/python3.12 "$CLI" run --registration "$REG" --output "$P-stop" --cycles 3 --stop-at-tick 5
# Public invalid-send probe: records unchanged metadata and enumerated carry/RNG digests; expect cycle1.
/usr/bin/python3.12 "$CLI" run --registration "$REG" --output "$P-invalid" --cycles 1 --reject-at-tick 5
For external live control from another CPU process:
/usr/bin/python3.12 "$CLI" status --output "$P-S1"
/usr/bin/python3.12 "$CLI" queue --output "$P-S1" x_bias-24mm
/usr/bin/python3.12 "$CLI" stop --output "$P-S1"
Do not inject external commands into frozen matrix evidence runs; use a separately reserved run for custom live-command evidence. For independent-copy S1 reproduction, change REPO to that copy and use another never-used output, same registration. Source hashes do not embed the development repository path; actual runtime binding captures the current copy's absolute paths.
Limits, evidence and dependencies
- Whole task shared ledger pre-reserves 2 Kit launches/run and
4500*requested_cycles+250physics calls, never refunds failed/unrun slots; maximum 24 launches/650000 calls. It holds the existing-task lock through cleanup to prevent parallel public runs. A fresh finite authority registration is required; its boottime deadline and the per-run deadline apply, never an unbounded grant. The historical 0.4.0 task registration expired2026-10-16 05:38:07 +08:00 and is not extended. The separate 0.4.1 GUI task records its own scope beginning2026-10-16 09:51:52 +08:00, with the same finite whole-task caps; its independent demo trials use fresh bounded sessions, additionally accounted together by the parent. Future user-invokedauthorizecreates a different bounded session without launching simulation. - Preparation: 360s total, zero physics. Physical run: ≤6 cycles1200s, >6 up to20 cycles2400s, stop probe900s; each includes120s cleanup. Calls: nominal20 ≤90250, six-cycle ≤27250, three-cycle stop ≤13750. One env/RTX5080, optimizer0.
- Existing process identity, fail-closed NVML, observed GPU diagnostics, stdout cap, native monitor, storage checks, authenticated worker FD, preclose same-object config binding, wait and owned cleanup are reused. Per-run owned cap4GiB and free headroom8GiB; shared ledger checks actual run/authority directory bytes against24GiB. Parent must additionally account for its package/analysis outputs outside those run directories. Native growth monitoring remains sampled, not a hard quota; existing coverage gaps remain reported.
- Automatic desktop admission uses only exact parent-approved executable PATH+SHA256 pairs in
configs/contact_cycle/skill_v4_desktop_allowlist.json; fresh PID/birth/argv/inode/hash captured each run, any compute or unapproved graphics rejects. Optional expert--gpu-identitysupplies an explicit current desktop identity JSON, never an old physical grant; existing GPUOwnership still rechecks every identity/client. - Results:
prepare/{binding,inventory}.json,preparation-reviewed.json, source snapshot, per-slot reservation/stdout/worker/supervisor results;run/{trace.pt,initial-state.pt,events-and-seams.json}, cycle audits, actual/requested counts and live receipts/pair history;mailbox/{status,receipt}.json; native/GPU observations; final waitedcampaign-result.json. Full500Hz samples and100Hz controls, integer local-cycle windows, no thinning/splicing/reset/reanchor. Evidence distinguishes physical functional gates and runtime/supervisor integrity. - Required same-layout repository files include all current
validatesource entries, original functional v2/v3 configs, frozen candidatelogs/contact_cycle/sim_residual_training_v1/prepared01/candidate_u0256.pt, its exactmanifest.jsonand complete134-filesource_snapshot/, all corresponding live frozen files/assets/configs, and the replay source hashes. Parent assembles this snapshot; do not regenerate its manifest. Candidate SHA256c088bb3aecaca751867118858626fd508f7a3983303802d7524353f80802cf99. - Installed runtime remains the pinned
/home/cen/IsaacLab,/home/cen/isaacsim-6.0.1, native numerical dependencies and RTX5080. Repository ROOT is relocatable; those installed dependencies are intentionally not made portable or re-pinned. The v4 runner explicitly setsSKILLFACTORY_GO2_ASSET_DIR=<this ROOT>/assets/generated/go2_flandSKILLFACTORY_LEG_BUNDLE=<this ROOT>/assets/generated/legs/go2_flbefore frozen import-time defaults run; retain the complete generated base Go2 FL, support and leg assets (including conversion.json/manifest.json/USD dependencies) and installed source pins. Seecontact_cycle_runtime.mdfor CPU namespace mapping and runtime constraints.
Historical 0.4.0 primary runs passed analytic20 continuous cycles, S1/S2/S3 six-cycle switches, u256 S1, requested stop at cycle1 and live invalid-command rejection, including preclose/GPU/exit/cleanup. Those are not a new-source/full-GUI matrix qualification for 0.4.1. See the separate v4 result/evidence index for independent-directory reproduction and precise acceptance; this interface document does not itself confer acceptance or general robustness/full-robot/real-world capability.