Skip to content

Simulation recordings

Real-model shared coordination

Recorded on 2026-10-11 with installed Robot Agent 0a91a2f0f0018aaadcb14587795a58ff2c9cb154, consuming immutable Harness 747d7d48a846ad361cf7260d2339b010c0b92a00. Both runs explicitly select the experimental coordination assembly and eight-second Owner terminal query window. GPT-6.1 Sol/high proposals use Codex CLI 0.159.0 on the host; the retained Ubuntu 22.04/Humble/Nav2/Gazebo amd64 scene stays offline at 2 CPUs/4 GiB. The model receives public map measurements and task context, without Session, execution authority or independent simulator truth.

Separate same-run clip Observed result and qualification
Normal navigation, 72.5 s Three actual proposals; A releases before B admission. Independent arrival evaluation succeeded, errors A 0.204 m/B 0.198 m. Final B remains pending/current. Original speed, followed by a two-second final-frame hold
Waiting-task withdrawal, 60.2 s Two complete proposals; the real final CLI turn is pending when the identified business Task receives SIGINT. Original B cancel/query and Session close, no final answer or new submit. B remains pending/revoked

Each 1600×900 recording uses a passive RViz observer in the actual scene. Startup is trimmed just before the first decision, model waits remain, and there are no internal cuts or speed changes. Captions use approximate same-run event times; the normal final-frame hold reports the recorded final state, not extra physical execution. Original/exported clips fully decode, representative frames are inspected and a focused independent check covers admission/withdrawal ordering and local process exit. The separate runs are not spliced together.

Task verdicts remain unassessed; the separate normal evaluator owns physical success. The withdrawal scene is not given a physical-success or hard-stop verdict. Native ROS cleanup reported shutdown-context errors, so complete native closure remains unknown. Remote inference cleanup is also unknown; local child/ process-group and container absence does not promote either. This sample is withdrawal after B's result while final assessment waits, not stopping in motion, outer-launcher graceful shutdown, hardware safety or broad model reliability. The earlier no-proxy model timeout and first evaluation import error remain recorded failures, not passed samples. Host deployment must separately verify its network path with TLS validation enabled.

Reproduce with the opt-in host-model tutorial. The withdrawal observer is a research qualification tool; the tutorial's controlled final-decision-stop cannot be combined with host-codex. Agent qualification records the corresponding source and limits. Existing videos below retain their own versions; none is relabeled as the newer model run.

Shared Agent coordination

These three separate qualified recordings use the shared private Agent coordinator with distinct navigation and manipulation contracts. Task decisions are controlled in all three; ACT supplies actual learned manipulation targets. They are not one continuous run or a real-model decision qualification. Existing public task defaults remain unchanged.

Clip Recorded sources Scope
Navigation, 41.4s Agent runtime 930eaf8273772b145f290b407a372950eed678f2, delivered in PR36 / 96edf715b41e41ac3127ef4a88b214153d4a31c3; consumed Harness 747d7d48a846ad361cf7260d2339b010c0b92a00 Public experimental-coordination normal A→B; separate physical evaluation succeeded, A0.207816m/B0.183400m. Final B pending/current; native cleanup unknown
Waiting-decision withdrawal, 39.7s Same Agent/Harness pair, a separate actual public-entry run After B native success, final controlled decision is held. Original B cancel/query and local Session close precede late-answer release; no new submit. Final B pending/revoked, native cleanup unknown
ACT transfer and hold, 9.02s Agent f708afa1c4577d2b8266ddc2ad0ec34e3edbfd93 / consumed Harness 4b4bc5bf28c249016cfe4d7e272fa9877766a838 Earlier private ALOHA assembly, MuJoCo, actual ACT400-step transfer then fresh feedback and a separate50-step hold. Both operations settled/released; separate one-second hold evaluation succeeded

Navigation was recorded on 2026-10-10 in the retained ROS Humble/Nav2/Gazebo amd64 image, offline at 2 CPUs/4 GiB with an explicit eight-second Owner terminal query window. A passive same-container RViz observer recorded these exact public runs; it did not drive decisions. Startup is trimmed near first submit, without internal cuts or speed changes. The1600×900 clips contain413/396 decoded frames at nominal10fps; screen-capture startup uncertainty and dropped frames remain, so these clips are not timing instruments. Withdrawal ordering is established by same-run trace and a focused independent safety check, not visually inferred from a stationary robot. This sample does not demonstrate stopping during motion, a hard stop deadline, remote/native cleanup or hardware safety.

ALOHA was recorded on 2026-10-09 in its retained MuJoCo/ACT environment, offline at6CPU/6GiB. The complete451-frame,640×480 episode plays at50fps in simulation time; inference/decision waits are omitted. It retains its original separately qualified version and evidence, not the newer public navigation tutorial's qualification. Task reports remain unassessed; the separate domain evaluators own physical-success verdicts. No old physics evaluator was rerun for publication.

Reproduce navigation using the explicit coordination tutorial. See Agent qualification and the private assembly design.

In-motion business revision

Two new recordings use installed Agent 189145996d2c7811ecd5d70641bc01c3e575a3d5 and its immutable Harness 7279cd12ae48d01384e081757c54c38b7858557b dependency, explicitly selecting scoped-two-context-nav2-revision-v1. The retained Ubuntu 22.04/Humble/Nav2/Gazebo amd64 image runs without network, at 2CPU/4GiB. No physical robot is used.

Both business instructions are controlled and delivered once after the single motion window opens: stop after four seconds, redirect immediately. Each scene has its own one-time physical evaluation; each clip contains only its own run.

Clip Proposal source Recorded outcome
Stop during motion Controlled proposal Exact A cancelled/no output, native closure/quiet/six B readiness replies, then settled/released in about5.050s. No B or completed site; stopped_by_instruction
Redirect during motion Three actual host proposals; explicitly requested gpt-6-sol / high via Codex0.159.0 A cancelled/no output and released in about4.994s before after-revision proposal/fresh B admission. Only B completed, independent error0.236m; final B current/pending

The model sees the bound redirect instruction and public map measurements, without a robot interface or independent Gazebo truth. Cancellation precedes the model wait; the model cannot grant reuse. Native/child/BT/outlet closure, fresh odometry quiet and successor readiness establish the reuse boundary. Passive Gazebo snapshots corroborate selected motion and B position, not an independent continuous quiet window. Prepared B resources can exist even when no B request is sent. Task verdicts stay unassessed and native/remote cleanup unknown. Exact scene/export containers and owned model/relay children were removed/reaped; host groups exited zero without forced cleanup.

Startup is trimmed, the RViz view cropped, captions added and the last frame held two seconds. The stop clip ends after task return; its later scene teardown is excluded. The redirect clip retains the model waits and task completion. There are no internal cuts or speed changes. Event captions use same-run monotonic time with screen-capture startup uncertainty; these videos and individual closure times are not timing guarantees. Old A paths may remain visible after cancellation and do not indicate live authority.

Separate no-instruction/foreign-identity/global-stop/actual B-pause scenes cover their selected outcomes. A first model scene failed in RViz startup before any proposal and remains incomplete; later unchanged runs passed without resolving that graphics failure's root cause. A separately evaluated four-second-delay model run completed the task but its RViz view stayed empty; that recording was rejected for presentation and is not spliced into these clips. Its graphics issue also remains unresolved. No arbitrary-route, Owner recovery, human intent recognition, hardware or general model-service qualification is implied. See the revision tutorial and qualification scope.

Checkpoint navigation

Two new same-run recordings exercise Robot Agent's explicit checkpoint task with the existing fixed Harness 13bc75e903f6005da3dd5d969f8242ce211d182f dependency. Recorded application code is 2ac951b0cafedf5e4871e689dbd5d6ec4d7e749c; new Linux Core and Agent installations use an existing qualified Ubuntu 22.04/Humble amd64 image. The scene has no network, 2 CPU and 4 GiB. No physical robot is used.

Clip Business and proposal sources Recorded outcome
Finish at A, about 28 seconds Controlled business instruction, delivered once after a four-second wait; controlled proposal A settled/released, then this new task completes at A; no B request. Separate same-run Gazebo evaluation measured A error 0.189m
Continue to B, about 68 seconds Controlled business instruction, delivered once after a four-second wait; three actual host gpt-6-sol / high proposals via Codex0.159.0 A releases before a single B admission with fresh feedback; final B current/pending. Separate same-run evaluation measured A/B errors 0.188/0.175m

The model receives the bound business instruction and declared map measurements; it does not choose business intent or receive physical-evaluator truth or a robot interface. Actual host model and relay children were reaped, both host groups exited zero without forced cleanup, and the exact scene containers were removed. Task verdicts remain unassessed; native and remote model cleanup stay unknown. Prepared B resources may exist even when no B request is sent.

Startup is trimmed, the RViz view cropped, captions added and the last frame held two seconds. Motion plays at original speed; caption alignment uses same-run monotonic events and retains screen-capture startup uncertainty. These clips are not timing instruments, in-motion replanning or hard stop evidence. Selected foreign and late instruction rejection runs are qualified separately; they are not silently inserted into these normal recordings. See the checkpoint tutorial.

Watch on the project website. The gallery contains actual ALOHA / MuJoCo and TurtleBot3 / Gazebo Classic recordings. Each silent clip identifies model-guided or deterministic decisions. Video shows motion; the corresponding observations and receipts establish execution and settlement. It does not establish a physical stopping guarantee.

Continuous ALOHA transfer and hold

Complete recording · Same-run summary · Runtime setup and caller

Recorded on 2026-10-01, using the explicit mujoco-stepped-handoff-v1 profile, seed0 and a deterministic caller. The external ACT worker supplies 400 transfer targets. After delivered settlement, the caller obtains fresh RGB/joints and requests a separate50-step hold of the last actually executed target. There is no reset between operations and no policy prediction during hold.

Same-run observation Value
Episode / observation sequence Epoch0 throughout; 0 → 400 → 450
Core operations IDs1/2; 400/50 returned steps; both outputs accepted and settled
Core domain verdict Unassessed; execution settlement is not task success
Separate task evaluation aloha-left-handoff-hold-v1: succeeded in the bounded one-second hold
Recording 451 frames, 640×480, 50fps, 9.02s; silent, uncut original

The video is copied from the final qualified m5d-handoff-session-02 episode. It follows simulation time (0.02s per step), not elapsed runtime: inference and caller waits have no frames. There is no crop, resize, speed change, internal cut or authored motion. The poster is frame350 from this same recording. All450 recorded steps matched the native continuation baseline, and all451 frames decoded. The linked summary selects existing report/evaluator fields and omits local paths; it is not the full raw trace or a second task evaluation.

Qualification used a development tree based on 53e65c4 with the continuous profile patch delivered with these assets. This is locally qualified on macOS arm64/MPS; see the backend version and dependency scope. Reproduction requires the external worker and migrated checkpoint described there; they are not installed or shipped with the Core package. Linux ACT/render qualification is recorded separately; this Mac video does not establish it. This single-seed demonstration is not a reliability estimate, an autonomous business Agent, self-improvement, or a physical grasp/stop guarantee. It does not demonstrate the separate cancellation checks. Raw experiment records and unsuccessful runs remain retained separately.

The clips below show TurtleBot3 / Gazebo Classic through RViz on Ubuntu 22.04 / ROS2 Humble. Their checker reads Gazebo poses and native/Owner observations.

Moving handoff

Readable excerpt · Full uncut recording · Owner events · Scenario verification

A starts toward (0.7, -0.5). After about 0.5 m of observed movement, the caller cancels it. The same Owner observes native work/drive closure and fresh stillness, prepares and checks a new context, settles A and admits B toward (0, -0.5). Green is A's path; blue is B's distinct task-scoped path. RViz may retain A's old path after cancellation; a visible path is not active execution authority.

In this run A moved 0.555 m and B moved 1.179 m. Fifty old-A commands were rejected at the drive while B was moving. That isolation claim comes from the native observations and checker, not from what the camera can show.

Final observation Value
A native result / output Cancelled / not delivered
A settlement Settled after closure and replacement readiness
B Admitted, native result accepted, native work closed
B settlement / third admission Pending / refused; no C context was prepared

Cancellation without a successor

Readable excerpt · Full uncut recording · Owner events · Scenario verification

A moved 0.557 m. The Owner sent one exact-goal cancellation and independently observed native termination, work/drive closure and a fresh quiet-motion window. No output was delivered and no B goal was sent. Ten injected late commands were rejected. The final receipt remains revoked and unsettled, and a second admission is refused. A successful scenario check means these expected facts were observed; it does not label the cancelled navigation task successful.

Recording provenance and edits

Captured on 2026-09-27 using the packaged cancel-moving and replace-moving scenarios. Runtime sources match commit 4911d79; this delivery adds capture/finalization only. The local visual dependency image was reused with the changed capture script. Runtime source comparison found no code differences; this was not a fresh dependency rebuild. This does not replace hosted clean-source simulation evidence in the testing guide.

Recording Uncut duration Captured frames Public excerpt
Cancellation 34.4 s 164 From 22 s to the end
Moving handoff 56.6 s 255 From 22 s to the end

Full recordings retain the original 1600×900 capture and timestamps, remuxed to MP4 without re-encoding. Capture requests 5 fps; frame drops under load mean frame count is not a clock. Excerpts remove only the first 22 seconds of startup, crop the 800×450 region at (400,300), and resize it to 1280×720. There are no internal cuts, speed changes, invented movement or simulated replacement frames. Posters come from the corresponding excerpt (8 s cancellation, 25 s handoff). Use logs for event ordering; no frame-exact event/video synchronization is claimed. The native result, observed quiet and settlement remain distinct. Neither timing from these recordings nor the emulated Linux environment establishes a stop bound.

Reproduce

From a checkout containing the recording support:

python3 integrations/ros2/simulation/simulate.py build --visual
python3 integrations/ros2/simulation/simulate.py run cancel-moving --visual --output simulation-results/record-cancel-01
python3 integrations/ros2/simulation/simulate.py run replace-moving --visual --output simulation-results/record-replace-01

Use a new output directory each time. recording.mp4 is the uncut capture; run.json must report a passed run and completed container cleanup, and verification.json must pass the scenario checks. Failed or interrupted runs may still produce video. See the simulation tutorial for requirements and failure handling. A model or real robot is not needed. These runs do not validate Owner restart, full network partition, general route replanning or physical-robot safety.

Earlier normal navigation

Download the original MP4.

This is an uncut 23.8-second, 1600×900 recording of a TurtleBot3 simulation on Ubuntu 22.04 / ROS 2 Humble, captured on 2026-09-21. Gazebo simulates the robot; RViz displays the robot and map. The recording is a stream-copy remux of the original capture, without speed changes or authored movement. The poster is an actual frame from the same recording.

The experimental Harness caller admitted one navigation goal, submitted it to Nav2 and accepted its result. The run observed approximately 2.485 m of displacement. It was captured from the normal-observation development tree based on b0929ab, before the final native-result log event and the later sequential-settlement work.

This clip demonstrates normal navigation. It does not demonstrate A→B handoff, movement-time cancellation, durable recovery or physical-robot safety. The observation-only path kept settlement pending and denied a second admission. Current behavior and validation are documented in Testing and ROS integration. The current simulation tutorial provides a source build, bounded scenarios and an optional live viewer. This historical recording does not demonstrate those later capabilities; the cancellation and handoff recordings above cover their own later scope.

Adding a major feature recording

Each major user-visible feature should include an actual-run demonstration. The project homepage embeds a video player; setup and full-evidence links supplement it. Record the implementing application and runtime together, identify model or deterministic decisions, preserve pending/unknown states and label any trimming, speed changes or held frames. A demonstration is presentation evidence for its stated run, not a replacement for tests, physical evaluation or hardware safety. Small fixes do not require a new recording unless the displayed behavior changes.

Model-guided navigation (local candidate)

Watch in the homepage player · Complete task excerpt · Same-run summary · Owner setup · Agent application

Recorded on 2026-10-03 from a local Harness candidate based on a145364 and Agent candidate based on 5e8b63b. The corresponding runtime and application are now delivered in Harness 6f32578 and Agent 9516daf; this recording retains its original local candidate scope. In isolated Ubuntu 22.04/Humble amd64 Gazebo simulation, the installed business Agent consumed public NavigationSession feedback. Host Codex0.159.0 requested gpt-6-sol / high and supplied three actual text proposals: visit A, visit B, observed complete. Model input was task context and public navigation pose/time/health; external Gazebo truth and evaluation did not drive decisions. The model ran on the host; the 2 CPU/4 GiB, network-disabled simulation received no credentials. Each proposal remained bounded at 30 seconds; the owner was explicitly configured for 45 seconds of caller idle time.

The scene ran 107.642 seconds. A/B same-run physical distances were 0.192253/0.166768m under the existing separate evaluator. A settled/released before B admission; B output was accepted with settlement pending. Agent assessment was observed complete, while task verdict stayed unassessed and native cleanup unknown. Local connection close or container removal does not invent a native reuse certificate. This single normal run does not establish model reliability, model motion cancellation, owner recovery or hardware safety.

The H264 video is 1280×800, 10fps and 70.9 seconds, silent. Startup is trimmed; the RViz scene is cropped/resized, same-run event captions are added, and the final frame is held for two seconds. There are no internal cuts or speed changes; model waits remain visible. Caption timing uses approximate monotonic event alignment with recorder startup latency retained, not a stopping-time instrument. The poster is a frame four seconds after A admission from this same final video. The summary selects retained facts without private prompts or operator paths; it is not a second task evaluation.

A prior recording failed after B: the owner's default 10-second final-close wait expired while a roughly10.03-second model proposal was returning. Its needs-help/unknown report and clean recorder/container cleanup remain retained separately. An explicit bounded caller wait resolves that budget incompatibility; no native deadline, freshness TTL, physical predicate or model budget was widened. Setup uses the paired commits above; the Agent manifest pins Harness 6f32578. Their installed combination checks preserve ALOHA and navigation consumption, without repeating this physical/model qualification. See Testing.

Presentation layout

The project README features the model-guided navigation recording through a GitHub video attachment, so readers can play it inline. It is the current local candidate with the most complete observation, decision and execution loop. The website keeps a featured player, three smaller players for ACT transfer, moving handoff and cancellation, and a separate two-player row for business stop and final proposal expiry. These recordings remain separate: their callers, time scales and qualification scopes differ. Original MP4 files and evidence remain available here; attachment URLs are presentation copies, not evidence IDs.

Business CLI interruption

Homepage player · Recording · Same-run summary

Recorded on 2026-10-03 with installed Harness 6f32578 and Agent runtime 9516daf. Controlled executable proposals visit A and B; after A releases and B is accepted and moves over 0.5 m, the public Agent CLI receives one SIGINT. The Agent explicitly cancels exactly B, without another proposal or admission. Core revocation, native cancel scheduling, a scoped drive seal/wheel-zero ACK, old-command rejection and a separate quiet/physical window were checked once in the same-run evaluation. The fixed physical window had 36 fresh samples over 3.502 simulated seconds, with endpoint drift about 0.000083 m and 0.00247 rad.

The report stays cancelled/task unassessed/native cleanup unknown/settlement pending. Explicit cancel response is retained; subsequent status/close connection errors remain visible in the task report. Proposal children and CLI were reaped; the scene container exited 0 without OOM and was removed. An earlier recording failed native preparation before any admission; that failure is retained separately.

The 1280×800 H264 recording is 30.9 seconds, original speed, with startup and view trimmed, approximate event captions and a two-second final hold. The poster is from B motion in the same export. It is not a stopping-time instrument.

Final proposal expiry

Homepage player · Recording · Same-run summary

Recorded on 2026-10-03 with the owner fix now delivered in Harness ea7b1eb and unchanged Agent runtime 9516daf. Controlled executable proposals visit A and B. The final child exceeds the original 30-second deadline, then emits a valid, matching answer while terminating and exits 0. The expired answer is not accepted; the Agent requests help and cancels exactly retained B. Native succeeded/output accepted remains alongside authority revoked/settlement pending; task verdict remains unassessed and native cleanup unknown. Status/close errors are retained. The Agent exits 1; owner/container exit 0, with children reaped and container removed without OOM.

One same-run evaluator confirmed A/B physical distances 0.198780/0.159009 m and A release before B admission. The first run exposed final cancellation being misclassified as an authority error; it is retained separately, without reevaluation. The focused regression failed before the fix and all 12 owner checks passed afterwards; completed normal close retains its prior wait path.

The 1280×800 H264 recording is 71.7 seconds at original speed, including the full proposal wait. Startup/view trimming, approximate event captions and a two-second final hold are declared; there are no internal cuts or speed changes. The poster is from B motion in this export.

Delivery and limits

Harness ea7b1eb1a075cd91ef590475f2496a48d19ca52a and Agent ffad065ae81ca9e9b17cba558d471078d657a9b2 deliver the matching fixed runtime and nine installed combination cases. Agent runtime source is unchanged; pairing and presentation do not repeat either recording's physical qualification. These clips use controlled proposals and a live owner, not real model failures. They do not establish complete native cleanup, Owner loss, restart recovery, hard stop deadlines or hardware safety. Both final exports decoded completely and their B/wait/final frames were inspected. Summary files select existing facts without operator paths or prompts; they are presentation metadata, not another evaluation. See Testing and the fixed Agent checks.

Finite Owner motion permission

Recording · Selected same-run facts · Explicit profile and reproduction

The installed product candidate uses scoped-two-context-nav2-owner-permit-v1, based on f43f4c3 with the corrected Owner patch delivered alongside these assets. The final installation passed21 focused Owner/Core checks and a new normal A→B run. In one isolated Ubuntu 22.04/Humble/Gazebo run, A settles/releases, B starts moving, and the actual Owner is killed with SIGKILL. Native command production continues; the drive independently expires the issue-stamped permission and actually writes every wheel target to zero. Late renewal and a fresh generation are refused. The separate evaluator confirms fresh external quiet and the fixed Gazebo snapshot window, declared to start 1.5 seconds after fault injection.

This run observed 1.013 seconds from the fault to zero application and 29.3 ms from the last permission deadline to application. These are observations, not hard time bounds. Native cleanup and Core settlement remain unknown, task success unassessed; there is no automatic restart or hardware qualification. A separate installed normal A→B and actual SIGSTOP check also passed their declared predicates; this recording demonstrates the SIGKILL case only.

The 1280×800 recording is about31.1 seconds at original speed. Startup/view are trimmed, approximate monotonic event captions are added, and the last frame is held two seconds. Display refresh is5fps; the recorder samples10fps. Full decoding and final-frame inspection passed. Captions are not a timing instrument. The poster is the final frame from this export. The JSON selects the existing unique evaluation; it does not rerun physical adjudication or contain local paths.

Single failure recovery

Homepage players · Normal A · Single backup B · Normal facts · Backup facts

Two separate 2026-10-05 isolated Humble/Nav2 runs use actual host-model proposals and the installed recovery business Task/public Session. Agent 682ead6247366f6c2507a58eb228322e7bf3e43f delivers runtime candidate 3bfa44b5efdd2aeec8c4dc2463e655ba22ba9415 (subsequent changes are documentation), paired with Harness 9278fea6249e61c1533defd4b39ea3d776f1a2ae. The requested model was gpt-6-sol, reasoning high, through Codex CLI 0.159.0. Private host logs/credentials remain outside container mounts; proposal answers cross the shared exchange. These recordings used private map preparation and map observation. They do not record the subsequently added public scene command.

Normal A uses two proposals, arrives within 0.209177 m and never submits B. In the backup run, A is actually accepted and returns status6 with no output on the static occupied-A planner map. Actual planner/BT closure, the sealed never-submitted controller, zero/quiet drive and six fresh B readiness facts precede Core settlement/release. Only then does one failure proposal select B; fresh identity/observation checks precede its admission. Three proposals are used; B arrives within 0.230241 m. A single evaluator qualified each run and one independent safety-decision review checked the ordering. The JSON files select those existing facts and are presentation metadata, not another evaluation.

The 1280×800 recordings are approximately31.1/48.8 seconds at original speed. Startup/teardown are trimmed, approximate event captions are added and the final frame is held two seconds; no internal cuts or speed changes. Both exports were fully decoded and start/branch/final frames inspected. Posters are their final frames. Captions are not timing instruments. Static map occupancy does not establish physical obstacles or general unreachability. Final B settlement is pending, task verdict unassessed and native cleanup unknown. These runs do not qualify Owner restart, hard stop bounds or hardware safety.