v0.8.0

Portable outcome reports and sample exchange

• Ship the paired policy outcome explorer in the installed package and complete
  offline Stress Lab. Retain gains and losses, all seed groups and the full
  experiment denominator. Export and independently verify the complete report.

• Include robot-reel-paired-outcomes.json as a checksummed release asset.
  Generate it with the installed wheel outside the checkout, reject an altered
  report, and compare the offline browser's filtered-view export with those
  exact bytes at desktop and mobile sizes.

• Publish the descriptive 30-trial / 20-pair result tables on Hugging Face with
  a source manifest and a reproducible export script. The recorded experiment
  is unchanged; this release adds inspection and distribution tools.

• Preserve Space homepage symbols and the Chinese link with HTML entities after
  public readback caught a UTF-8 character damaged at the hosting boundary.

• Connect the Space and both READMEs to a curated Hugging Face model/data
  collection and the public feedback thread. Credit the exact recording sources.

• Verify uploads anonymously against the tested CI artifact: every Git/LFS
  object plus the public manifest, homepage, three viewers and thumbnail.
  Bound CDN retries, reject stale or changed deployments and preserve unrelated
  remote files.

• Host Cloth, Stress and Butterfly labs as a native Hugging Face static Space,
  with original recordings, a dedicated landing page and model/dataset credits.
  Publish the exact browser-tested artifact after the current main commit's
  checks pass; verify its full inventory and source identity before upload.

• Make sample links visible and directly openable when embedded browsers deny
  clipboard access. Preserve the selected frame, condition and camera.

• Open Cloth Lab sample JSON files, verify their source facts and restore their
  sample and camera offline. Reject changed or ambiguous files without moving
  the current view. Add cloth --verify-sample for independent standard-library
  checks against a verified source bundle.

• Export Cloth Lab samples as 1920 × 1080 PNG figures with original mesh geometry,
  measured deformation, source clocks and a positions fingerprint, plus portable
  JSON records with full-precision metrics. Both work offline. Shared links now
  preserve camera rotation; failed PNG encoding can be retried.

• Point the homepage, both READMEs and the published Stress viewer at the
  verified 0.7.1 downloads, installation guide and checksums.

Start with START-HERE.md. Extract either experiment ZIP and open its index.html
locally. Cloth Lab includes all 42,471 recorded vertex samples, original velocities,
the editable USD and native readback reports. The standalone cloth USD contains
the same bytes as the scene in the ZIP; import it into Blender at 30 fps.
For Stress Lab, import robot-reel-seed-09-review.json to inspect a selected moment
in the complete thirty-trial experiment. Both ZIPs are the exact artifacts
exported by the installed wheel and tested offline in Chromium at desktop and
mobile sizes. No GPU or Python is needed to use these recorded browser labs.

Stress Lab includes all four paired outcome groups and complete report export.
robot-reel-paired-outcomes.json was exported and verified by the installed CLI;
the offline browser produces the same report even after selecting a subset.
Use robot-reel stress stress-lab --paired-report robot-reel-paired-outcomes.json
to check it independently. Cloth Lab includes source-checked sample import,
portable JSON and 1080p figure export.

The wheel, source distribution and Rerun recording passed the same six validation
jobs. Open robot-reel-seed-09.rrd in Rerun 0.37.2. SHA256SUMS covers every asset.
The original recordings are unchanged; neither policy inference nor cloth
simulation was rerun for this release. PyPI publishing is configured separately.

v0.7.1

Reliable offline export recovery

• Stage Stress Lab exports beside their destination and publish only after media,
  telemetry, manifest and archive verification succeeds. Failed or interrupted
  exports can be retried without removing a partial output. Reject overlapping
  input/output paths, symlink destinations and occupied outputs; preserve files
  another writer creates before publication.
• Connect the homepage and both READMEs to the verified 0.7.0 release files,
  with direct downloads for both offline labs and the editable cloth scene.

Start with START-HERE.md. Extract either experiment ZIP and open its index.html
locally. Cloth Lab includes all 42,471 recorded vertex samples, original velocities,
the editable USD and native readback reports. The standalone cloth USD contains
the same bytes as the scene in the ZIP; import it into Blender at 30 fps.
For Stress Lab, import robot-reel-seed-09-review.json to inspect a selected moment
in the complete thirty-trial experiment. Both ZIPs are the exact artifacts
exported by the installed wheel and tested offline in Chromium at desktop and
mobile sizes. No GPU or Python is needed to use these recorded browser labs.

The wheel, source distribution and Rerun recording passed the same six validation
jobs. Open robot-reel-seed-09.rrd in Rerun 0.37.2. SHA256SUMS covers every asset.
The original recordings are unchanged; neither policy inference nor cloth
simulation was rerun for this release. PyPI publishing is configured separately.

v0.7.0

GPU cloth and portable deformable experiments

• Ship Cloth Lab in the installed package and release downloads. Export a checked
  recording and its complete offline ZIP without a source checkout, Newton or a
  GPU. Preserve the original data and native reports; request native USD readback
  separately. Package the method and license, protect existing input/output
  directories, and publish only a fully validated export.

• Test both installed-wheel archives offline at desktop and mobile sizes, and
  publish those exact files with checksums. Keep citation, package and release
  versions aligned. Existing recordings and earlier release assets are preserved.

• Add Cloth Lab: three independent Newton VBD cloth simulations recorded on
  NVIDIA L40S, with original float32 positions and velocities, shared-clock
  mesh comparison, orbit/overlay replay and a portable offline experiment.
  Check every deforming vertex and velocity through OpenUSD, and every imported
  mesh vertex through Blender. Publish a source-mapped preview and document
  the numerical settings, diagnostics and material-validation limits.

• Point the homepage, READMEs and published Stress Lab at the verified 0.6.0
  offline bundle. Add a homepage download section with the sample review, guide
  and checksums after the release assets are publicly available.

Start with START-HERE.md. Extract either experiment ZIP and open its index.html
locally. Cloth Lab includes all 42,471 recorded vertex samples, original velocities,
the editable USD and native readback reports. The standalone cloth USD contains
the same bytes as the scene in the ZIP; import it into Blender at 30 fps.
For Stress Lab, import robot-reel-seed-09-review.json to inspect a selected moment
in the complete thirty-trial experiment. Both ZIPs are the exact artifacts
exported by the installed wheel and tested offline in Chromium at desktop and
mobile sizes. No GPU or Python is needed to use these recorded browser labs.

The wheel, source distribution and Rerun recording passed the same six validation
jobs. Open robot-reel-seed-09.rrd in Rerun 0.37.2. SHA256SUMS covers every asset.
The original recordings are unchanged; neither policy inference nor cloth
simulation was rerun for this release. PyPI publishing is configured separately.

v0.6.0

Portable reviews and a complete offline lab

• Deliver the complete Stress Lab ZIP, a source-checked sample review and a
  start guide alongside the wheel, source distribution and native Rerun recording.
  Publish the exact archive exported by the installed wheel and tested offline
  in Chromium. Keep these assets separate from the Python distributions and
  include every downloadable file in SHA256SUMS.

• Export selected Stress Lab moments as portable JSON and readable Markdown with
  separate user notes. Import and compare facts before restoring a selection;
  verify reviews against a complete local collection with the standard-library
  CLI. Preserve source clocks, held/final observations, active inference records,
  checkpoint identity and the complete experiment's counts.

• Guide first-time visitors from a selected paired outcome to native inspection
  and full-experiment reproduction. Add purpose filters with shareable URLs,
  keyboard controls, browser history and offline/no-JavaScript fallbacks.

• Load homepage motion only after an explicit play action. Preserve the original
  Butterfly preview's 61 source samples and clock in a checked H.264 derivative;
  pause when the preview leaves view, the tab is hidden or reduced motion changes.

• Synchronize the complete static landing template, including its markup and
  styles, while retaining script-only synchronization for recorded replay pages.

Start with START-HERE.md. Extract robot-reel-stress-experiment.zip and open
index.html locally; import robot-reel-seed-09-review.json to inspect a selected
moment. The ZIP is the exact artifact exported by the installed wheel, verified
against all thirty trials and tested offline in Chromium. No GPU or Python is
needed to use the recorded browser lab.

The wheel, source distribution and Rerun recording passed the same six validation
jobs. Open robot-reel-seed-09.rrd in Rerun 0.37.2. SHA256SUMS covers every asset.
The original trials are unchanged; new policy inference was not performed for
this release. PyPI publishing is configured separately.

v0.5.0

Paired policy inspection and verified distributions

• Bundle Stress export notices, license and methods as package resources. Build
  a wheel from the source distribution, install it in a fresh environment, and
  export all thirty trials outside the checkout, checking videos, MCAP records
  and the complete offline archive.

• Gate Pages and tagged releases on the same six validation jobs. Require
  matching tag, package and changelog versions; publish the tested distributions
  and native Rerun recording with SHA-256 checksums. PyPI publishing is opt-in
  after trusted-publisher setup.

• Run the Docker image as an unprivileged user with a regular package install.
  Check the default command, writable output, and an actual storyboard export
  through a bind mount using a different UID/GID.

• Add a native Rerun inspection workspace: select a paired seed, open six embedded
  videos alongside measured 3D paths, applied controls and inference timings, or
  download the portable RRD. Read back native poses, controls, source JSON, video
  byte streams and clocks; retain terminal semantics and the full experiment's
  denominator. Feature the verified seed-09 recording on the landing page and in
  both READMEs. Rerun 0.37.2 remains an isolated optional dependency.

• Synchronize replay scripts, matching offline ZIP members and dependent manifest
  hashes together. Integration tests update real VLA, Chaos and Stress packs,
  verify their evidence and archives, and check that repeated syncs change nothing.

• Keep the Stress Lab's offline-archive link in its template. Custom builds
  default to their own experiment.zip; --archive-href explicitly selects the
  published Release asset. Tests cover the complete published page, local build
  downloads and CLI overrides.

• Encode every recorded video at one quality. Two recorders had drifted to
  imageio-ffmpeg quality 9 (crf 5) and wrote the heaviest media in the
  repository; a test now holds all eight writers to the same value.

• Hold every published page inside a compressed transfer budget, measured the
  way GitHub Pages serves them. The Stress Lab page is 1.26 MiB of the 1.50 MiB
  limit.

• Document opening the Stress Lab's MCAP telemetry in Foxglove, with FoxQL
  expressions for paired conditions, controls and inference timing, and record
  what Rerun decodes today. A test resolves every documented expression and
  message count against the real telemetry.

• Ship the Stress Lab's 56 MB offline archive as a release asset instead of a
  tracked file, verify a published site without it, and fail the tests when any
  tracked file exceeds 25 MB. The archive was removed from the repository
  history as well, so existing clones must be re-cloned or hard-reset.

• Check that every relative link and heading anchor in the Markdown docs and the
  landing page resolves, with the standard library.

• Add launch copy for the landing page, Stress Lab and Butterfly Lab to the
  maintainer notes.

• Publish a landing page at the site root that indexes all twelve recorded demos,
  with a stdlib quick start and a copy button; the Microduck replay keeps its
  own page. The landing template lives in scripts/landing.html and is covered
  by the viewer script check.

• Move the quick start to the top of both READMEs, add live-demo, Python and
  stars badges, a Colab notebook and a Dockerfile for the stdlib checks, and a
  section on how Robot Reel differs from LeRobot, MuJoCo Playground, Isaac Lab,
  Genesis, Newton, Rerun and Foxglove.

• Add a 1280×640 social preview built from the Butterfly Lab poster, issue and
  pull request templates, a code of conduct, a security policy, a citation file,
  an examples/ guide and a tag-triggered release workflow that builds the
  package and publishes it to PyPI through trusted publishing.

• Move maintainer launch notes out of the published docs/ tree into notes/.

• Type-check each replay template's inline viewer script, and check the published
  pages still carry it. Copying a template into its pages re-records the affected
  bundle hashes along the manifest chain.

• Add the Stress Lab: thirty closed-loop SmolVLA CUDA trials on NVIDIA L40S across ten paired
  initial states and three native lighting/camera conditions. Preserve the
  locked plan, every result and interruption history, input/noise hashes,
  initial physics and separate policy/simulation timings.

• Add synchronized paired replay, an all-trial outcome matrix, source-aware
  final-frame holding, inference/outcome navigation and measured end-effector
  separation on the shared recording prefix.

• Export the complete offline experiment, CSV and lossless JSON telemetry in
  MCAP; verify every message on readback, all sixty camera videos and published
  summary/HTML against the source traces.

• Add a source-mapped homepage preview and a dated Chinese research roadmap.

• Add the Butterfly Lab: twelve isolated CPU Newton worlds with a 0.05° release
  sweep, interactive motion overlays and time sculptures, source-derived
  separation metrics, a portable offline bundle and a native-checked OpenUSD
  animation retaining all 14,424 body samples.

• Feature the new experiment in both GitHub READMEs with a source-mapped GIF
  and a reduced-motion poster.

• Redesign the English and Chinese GitHub homepages around a recorded-motion
  cover, a reduced-motion fallback, scene cards and direct demo/download links.

• Add Physics → Cinema: an interactive divider between the original MuJoCo
  comparison and its Blender replay, with shared sample stepping, contact
  navigation and source consistency checks.

• Preserve detailed recording and development instructions in dedicated guides.

The wheel, source distribution and native Rerun recording below passed the same six validation jobs as the site. Open robot-reel-seed-09.rrd in Rerun 0.37.2; its six videos and original evidence are embedded. PyPI publishing is configured separately.

v0.4.0 — SmolVLA tasks and an MCP Blender director

Robot Reel 0.4.0 makes Physical AI runs easier to watch, inspect and reuse.

• SmolVLA task replay: an actual CPU policy rollout in LIBERO, with synchronized scene/wrist cameras, every applied action, measured robot states and the task outcome. The published seeded run completed the bowl-to-plate task after 76 actions.
• MCP agent director: inspect recorded events, create a validated storyboard, and build an editable Blender film with four cameras, captions and sample-preserving slow motion. The included film retains 180 source samples across 210 frames.
• Newton → OpenUSD → Blender: the existing CPU rigid-body replay and native transform checks are included in this release.

Try the demos:

• https://noteflowai.github.io/robot-reel/vla/
• https://noteflowai.github.io/robot-reel/director/
• https://noteflowai.github.io/robot-reel/newton/

Downloads below include the complete offline VLA episode and the editable director project. Preserve the media attribution included in the episode. Reproduction instructions and the Chinese README are in the repository.

Validation: 46 Python tests, 18 browser checks, actual MCP stdio calls, real CPU policy inference, and native Blender checks of all 420 directed vehicle samples. Videos and downloads are tied to their recorded traces by checksums. The VLA result is one simulation rollout, not a benchmark or hardware validation; the website plays recorded examples.

v0.3.0 — Two runs. One clock.

Two runs. One clock. Compare Microduck speed commands and early/late braking with synchronized raw footage, frame stepping, measured channels and downloadable evidence.

• Microduck demo: https://noteflowai.github.io/robot-reel/compare/microduck/
• Braking demo: https://noteflowai.github.io/robot-reel/compare/braking/
• Reproduce: https://github.com/noteflowai/robot-reel/blob/main/docs/comparison.md

New CLI: robot-reel compare --left …-trace.json --right …-trace.json --output ….
Microduck accepts --speed from 0 to 0.6 m/s. Comparisons reject mismatched times, lengths, model configurations or engine versions. Playback follows the left video clock with drift correction; pausing aligns both sides to one recorded frame.

Download a comparison MP4 to share, or extract a comparison bundle and open index.html directly for inspection. Bundles include original selected traces/raw videos, source manifests and checksums. The wheel includes both browser templates. The original three showcase films remain available in v0.2.0.

Validation: 22 Python tests and 10 Chromium browser tests passed locally. Both comparison films fully decoded; wheel template inclusion and phone-width layout were checked.

These are individual simulation trials, not a statistical benchmark. Microduck runs the official ONNX policy with XML PD actuators, not BAM or hardware. Automotive braking is a 1D scripted surrogate, not autonomous driving or a safety validation. Microduck model media retains upstream BY-SA-NC terms: keep MICRODUCK-MEDIA-NOTICE.txt with redistributed Microduck footage. Recorder code is Apache-2.0.

v0.2.0 — Microduck, braking, and inspectable physical AI

Robot Reel now records the official Microduck walking policy, a repeatable braking comparison, and editable arm shots—and exports a browser replay with every film.

• Microduck: official pinned ONNX policy at 50 Hz; 300 simulation frames, 14 joint channels, policy actions and model references. XML PD actuator approximation, not BAM or real hardware.
• Automotive: identical 8 m/s initial speed and obstacle, two scripted brake thresholds. Late braking records contact; early braking stops with an 8.661 m gap in this 1D surrogate. No driving foundation model was run.
• Interactive replay: shot navigation, single-frame stepping, joint/reference plots, timestamp links and local-file playback.
• Arm remixes: download a JSON plan from the browser and record it with --shots.
• Evidence checks: required hashes plus action/frame, policy/frame and contact-summary agreement.

Validation: 18 Python tests and 8 browser tests; real captures and full video decoding; cold-cache official asset download; wheel contents and installed template; existing-pack re-edit smoke test.

Microduck films are 15 seconds; braking films are 17 seconds; studio films are 31 seconds. Each evidence ZIP contains its local replay page, original simulator recordings, traces, provenance and license notices.

Microduck media depicts Pollen Robotics' assets, identified upstream as Creative Commons BY-SA-NC. Keep the included attribution and noncommercial/share-alike terms with that footage. Robot Reel's recorder code remains Apache-2.0.

Robot Reel v0.1.0 — Words in. Motion out.

Robot Reel turns a recorded MuJoCo simulation into a shareable film with inspectable motion traces.

• 31-second landscape and portrait videos with measured joint overlays and an original synthesized soundtrack.
• SO-100: four live-agent tool calls, 480 physics frames, final home error 0.008941 rad.
• G1: 300 explicitly labeled kinematic pose frames; no walking or balance-policy claim.
• Credential-free scripted demo, optional Bedrock director, bilingual setup and provenance checks.

Validation: local target checks and evidence-tampering test pass; both H.264 videos fully decoded; browser playback and mobile layout checked. Wheel build passes.

Download either MP4 to share, or the evidence ZIP for original simulator recordings, per-frame traces, prompt, model response, checksums, and model license notices. This is an early preview with a pinned Strands Robots backend adapter.