RobotReel
Physical AI replay lab

Give Physical AI
a replay button.

Play the motion. Inspect the failure. Keep the evidence behind every recorded experiment.

Inspect a failure ↗All experiments ↓

Recorded simulations · No account or install

Explore all 12 worlds ↗
Recorded Newton poses · preview at 3.33× speed · sculpture depth represents time. Motion starts when you press play.

GPU Cloth Lab ↗Explore the Hugging Face labs ↗Follow a guided review ↓

QWEN3.8 / MODEL EXPLANATION REVIEW

Read the explanation.
Check the recorded facts.

Six actual model outputs beside three original SmolVLA episodes. Compare images-only review with review supplied the outcome record, then inspect every claim offline.

Selected development examples · explanations are not verified causes

Review the model outputs ↗Download the full review ↓
Watch the 30-second first review

Four annotated views of the actual interface. No audio or new model run. Run the first review locally ↗ · Transcript · Source hashes

Start here · a guided review

From a surprising run to the evidence.

Follow one selected example, then check the full experiment yourself. No setup to explore.

  1. 01 / CompareSame seed. Different endings.Open seed 09 at its largest shared trajectory difference. Reference succeeds; reduced light reaches the step limit.
  2. 02 / InspectRead the run behind the result.Put the cameras, measured 3D paths and applied controls on one timeline in the Rerun workspace.
  3. 03 / ReproduceCheck all 30 trials locally.One task, 10 initial states, 3 conditions. Verify every recorded trial and keep the source data and checksums.
MICRODUCK / INSIDE A LEARNED WALK

A tiny robot.
Every joint on record.

Rotate a 3D joint schematic beside the original walk. Overlay policy targets, inspect the residual heatmap, and switch between two recorded speed commands.

14 joints · 8,400 measured samples · body-fixed schematic · simulation with PD actuators

Explore the Motion Lab ↗Export a frame. Verify it locally ↗Anatomy inspiration ↗
Microduck joint structure and original recorded video with measured versus target angles.
Original simulation and model-derived schematic. No source code or assets copied from the Anatomy Space. Model geometry and footage keep their upstream noncommercial/share-alike terms.
Native Rerun inspection

Same seed.
Different endings.

One run succeeds. Two reach the step limit. Put their camera views, measured 3D paths and applied controls on the same timeline, then inspect the policy calls behind the motion.

6 embedded videos · 405 observations · 41 policy calls
Open the Rerun workspace ↗Download recording · 4.3 MiB ↓

A selected group of 3 trials from the 30-trial Stress Lab. The full experiment remains available. Desktop browser recommended; the downloaded file opens locally in Rerun 0.37.2 with all six videos included.

Inspect this pair in the browser replay ↗ · Rebuild and verify

Native Rerun workspace showing three recorded policy cameras, measured end-effector paths in 3D, and the applied gripper-control curves.
Actual Rerun workspace · source seed 09. The 3D lines show complete recorded paths; camera titles state each run's outcome and end time. This selected example is a diagnostic, not a benchmark score.
Offline experiment downloads

Recorded labs. Ready to take away.

Inspect Microduck joints, compare simulation error, explore cloth deformation, or review paired policy trials. Each lab includes original data, a replay and verification tools.

Microduck: 8,400 joint samples · Solver: 366 states · Cloth: 42,471 vertex samples · Stress: 30 trials
Microduck Lab ↓Solver Lab ↓Cloth Lab ↓Stress Lab ↓Sample policy review ↓

Extract each ZIP into its own folder and open index.html. No account, Python or GPU is required to replay. The installed CLI checks source data and exported frames. Downloads below follow the latest release; its checksums identify the exact files.

Quick start guide ↗ · SHA-256 checksums ↓ · Installation packages & all files ↗

Quick start · Python 3.12+, standard library only

Check the experiment yourself.

Verify all 30 Stress Lab trials and the original SmolVLA episode, then turn the recorded braking comparison into a checked storyboard.

git clone https://github.com/noteflowai/robot-reel.git
cd robot-reel
# Check all 30 paired trials and the complete evidence.
python3 -m robot_reel.cli stress docs/stress
# Check the recorded policy episode and its evidence.
python3 -m robot_reel.cli vla docs/vla
# Turn the included braking comparison into a checked storyboard.
python3 -m robot_reel.cli direct docs/compare/braking \
  --plan examples/contact-storyboard.json --output artifacts/director

Install the released CLI from PyPI; see the installation and Docker guide. To collect new GPU trials, follow the recording recipe.

Choose an experiment for your workflow.

Showing all 19 recorded demos

Review policy outcomes in Stress Lab, inspect numerical error in Solver Lab, or edit captured assets in Blender Scene Lab. Each includes source data and a reproduction guide.

A real SO-101 LeRobot episode: two camera views on one clock above commanded versus measured joint curves.
LeRobotDataset v3 · SO-101 · your data

Your dataset.
One command.

Turn any LeRobotDataset episode, on the Hub or on disk, into an offline replay: every camera, commanded versus measured joints and fingerprinted source files.

 Microduck on captured coastal terrain, rendered in Blender from its recorded simulation camera.
Microduck · MuJoCo → Blender

One frame.
Back in the scene.

Follow the recorded robot, inspect terrain contacts, and match its original camera in Blender. Both runs, all 362 frames, editable native projects.

 SmolVLA baseline source cameras in the recorded LIBERO-Plus subset.
LIBERO-Plus conditions · one paired state

A changed view.
A different outcome.

One paired initial state, three native conditions. Baseline and lighting succeed; the camera variant reaches the step limit. Inspect every source frame and action.

 Cycles render of a procedural factory hall with CNC machines, a weld cell and cyan digital-twin status rings linked to an operations center.
Factory Twin · Blender 5.2 · OpenUSD

Same shift. Same failure. Close the loop.

A digital twin estimates hidden spindle wear, forecasts billed demand and acts on the plant. Compare it with the same twin in shadow mode across 12 paired shifts.

 Six recorded Genesis and Newton flights compared with an analytic trajectory and timestep error curves.
Genesis × Newton · L40S CUDA

One launch. Mind the timestep.

Change the timestep, inspect the shift, and compare every position and velocity with the analytic flight. Native replays, editable USD and the complete offline lab.

 Three real GPU cloth recordings with different bending coefficients, fixed vertices and source clocks.
Cloth Lab · L40S · Figures & OpenUSD

Same sheet. Three ways to fall.

Export a slide-ready figure, or open a shared sample and check it against the original recording. Orbit the meshes and take the scene into Blender.

 Three SmolVLA rollouts from one initial state under reference lighting, reduced light and a shifted camera.
Stress Lab · SmolVLA · CUDA

Same task. Change the view.

Inspect all 30 trials. Group paired successes, gains and losses; open any seed and export a complete outcome report for independent checks.

 Neon 3D time sculpture of twelve Newton pendulum trajectories.
Butterfly Lab · Newton · OpenUSD

0.05° apart. Worlds apart.

Twelve isolated worlds begin at nearly identical angles. Drag to orbit their recorded paths and find where a tiny release difference becomes a 6.26 m gap.

 Split view between the original MuJoCo footage and its Blender replay.
Physics → Cinema · MuJoCo · Blender

One recording. Two looks.

Drag between the original simulation and its Blender replay. Same source samples, same simulator clock, a different camera and scene.

 SmolVLA's scene and wrist cameras after the recorded bowl-to-plate task.
SmolVLA · LIBERO · 76 actions

Words → robot actions

A language instruction becomes an actual SmolVLA rollout. Inspect both camera views, measured state and each applied control.

 A Blender camera shows two recorded braking trials at the contact sequence.
Agent director · MCP · Blender

Brief → Blender film

An MCP agent plans camera cuts, captions and slow motion. Every frame of the editable film maps back to the original recording.

 Recorded Newton double-pendulum poses and their motion trail in the browser's 3D replay.
Newton · OpenUSD · 362 transforms

Physics → editable 3D

Record Newton physics on CPU, explore measured poses in the browser, then open the animated OpenUSD scene in Blender.

 Microduck Motion Lab: recorded video, interactive 3D joint schematic and exact measured versus target curves.
Microduck · 14 joints · source-checked

Inside a learned walk.

Orbit a joint schematic, overlay targets and inspect 8,400 measured joint samples. Compare two walks and take the original evidence offline.

 Microduck's official ONNX walking policy recorded in MuJoCo, with measured joint telemetry.
Microduck · ONNX policy · 50 Hz

A tiny robot. A learned policy.

Pollen Robotics' Microduck walks with its official policy. Follow 14 joint targets, measured responses and base motion frame by frame.

 Two Microduck runs at different speed commands, side by side on one clock.
Comparison · two runs · one clock

Two speed commands

Two independent Microduck captures aligned on simulation time. Step both views together and read the measured difference.

 Early and late braking trials side by side.
Comparison · braking · contact

Early vs. late braking

Two one-dimensional MuJoCo trials with the same initial conditions and different brake thresholds. Jump to the recorded contact.

 The recorded braking scene rebuilt as an editable Blender project.
Blender · editable scene

Keep the motion. Change the look.

Download the braking replay as a Blender project whose keyframes come straight from the recorded trace.

 A single recorded braking trial with its telemetry.
Braking · single run

One braking trial

The single-run replay behind the comparison: applied controls, measured state and frame links.
SO-100 arm · MuJoCo · studio

SO-100 arm studio

A recorded SO-100 arm scene with shot navigation and per-frame telemetry, the starting point for the example shot plans in the repository.

Keep the run behind the film

Evidence travels with every demo
Watch

Watch recorded behavior

Browser replays with synchronized views, shot navigation and frame links. The browser plays recorded simulations; nothing is re-simulated on your machine.

Inspect

Every applied action

Applied controls, measured poses, recorded outcomes, source revisions, hardware and timing. Validators check hashes, timestamps and frame mappings.

Reuse

Take the scene with you

Offline episodes, MP4s, JSON and MCAP traces, editable Blender projects and animated OpenUSD scenes.

RESEARCH EXAMPLE / FROM RECORDING TO AGENT EVIDENCE

A skill loaded. Did the task pass?

27 real GPU model trials connect Robot Reel recordings, Skills Anywhere delivery and EvalArc grading. All failures, MCP receipts and candidate files are retained.

Inspect the Skill Impact Lab ↗ OpenEnv: verify the native edit ↗
Review independent source tasks

A separate 36-attempt SWE workflow review uses three public Python projects and four fixed conditions. No attempt obtains acceptance: 31 have assessable reports and five remain uncertain. Inspect eight nonempty patches, source-path failures and native grader labels. These are coding-task records, separate from the robot-trajectory examples.