Give Physical AI
a replay button.
Play the motion. Inspect the failure. Keep the evidence behind every recorded experiment.
Recorded simulations · No account or install
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
Take the recorded experiments offline with release 0.19.2. Download the labs and independently verify their recorded evidence ↗
Choose an experiment for your workflow.
Review a policy, inspect a simulation, or edit a captured scene.
Start with paired policy outcomes, simulation error curves, or a captured scene edited in Blender. Download the corresponding source data and follow each lab's reproduction guide.
Same shift.
Close the loop.
A digital twin estimates hidden spindle wear, tests what-if plans and acts on a simulated plant. Compare 12 paired shifts with the same twin in shadow mode.
Open the Factory Twin ↗
A model explains.
What can you verify?
Compare six actual model outputs with three original robot episodes. Switch the supplied evidence and check concrete claims independently.
Open the model review ↗
One frame.
Back in the scene.
Step through the robot, recorded camera and terrain contacts. Compare matching source and Blender frames; download both animated native projects.
Open the Scene Lab ↗
A changed view.
A different outcome.
Three runs across baseline, camera and lighting conditions: two successes and one step limit. Compare source frames and every applied action.
Open the Plus Lab ↗
One launch.
Mind the timestep.
Six CUDA runs, three timesteps, one analytic reference. Inspect the error, export every sample, and reopen the native trajectories or editable USD.
Open the Solver Lab ↗
Inside a
learned walk.
Orbit a joint schematic, overlay policy targets and inspect every tracking residual. Two original walks, source-linked JSON/CSV and an offline lab.
Open the Microduck Motion Lab ↗
Same task.
Change the view.
All 30 trials, failures included. Group paired gains and losses, open the source recordings, and export the full outcome report.
Open the Stress Lab ↗
Same sheet.
Three ways to fall.
Orbit and overlay recorded meshes. Export a 1080p figure, verify a shared sample, and bring the original geometry into Blender.
Open the Cloth Lab ↗
0.05° apart.
Worlds apart.
Turn a small release difference into an inspectable motion sculpture. Follow every original pose and find the largest measured separation.
Open the Butterfly Lab ↗Keep the run behind the image.
Trajectory labs retain source clocks and recorded samples; Scene Lab preserves input assets, edit recipes and native readback. Download the evidence and follow each lab's verification guide.
Take the scene into Blender.
Cloth and pendulum OpenUSD exports retain recorded geometry and poses. The full project adds Blender films, agent-directed storyboards and native Rerun inspection.
Explore Physics → Cinema ↗Skill Impact Lab: inspect all 27 real GPU trials ↗
Related engineering audit: inspect 36 independent-source SWE attempts, including 31 assessable reports and five uncertain outcomes. This separate coding study links MCP delivery, exported patches and native grading; no attempt obtained acceptance.
Methods and reuse
Results describe the recorded task and simulation conditions. Each lab includes its method and source data for review.
What the experiments cover
- Stress Lab and LIBERO-Plus use separate conditions and action budgets. Full benchmark evaluation and real-robot performance require separate runs.
- Cloth coefficients are solver settings, with uncalibrated fabric properties. Collisions and self-contact are disabled.
- Microduck Motion Lab uses a body-fixed schematic; Scene Lab preserves recorded floating-root motion. Both use simulated PD actuators.
Microduck model-derived geometry and footage retain noncommercial/share-alike terms. See the policy media notice and source commit and file checksums before reuse.