Inside a
learned walk.
Pick a joint. Follow what the policy asked for and what the robot did in simulation. Two recorded walks, one inspectable clock.
Joint structure / 3D schematic
Original recording
Loading the original recording…
Playback controls ↑Every joint. Every frame.
|measured − target| · 0 → 0.5+ radKeep the evidence.
- Video time
- Simulation sample time
- Preceding policy step
- Forward command now
- Measured speed now
- Mean speed while commanded
- World base X / Y
Commanded speed and measured speed differ. Body attitude is visible in the original video; the schematic fixes the floating root because its orientation was not saved.
Open a shared frame JSON to check its recorded facts and restore the run, frame and joint. Your orbit view stays in place. The file stays in this browser.
The ZIP contains both original videos, full policy traces and this viewer. Unzip and open index.html. No live inference, external runtime downloads or server required.
How the schematic is built, and what it cannot tell you
Forward kinematics uses the pinned Pollen Robotics model and the original 14 measured or target angles. Each box bounds one body's visual geometry; it is an abstraction, not a CAD mesh or collision model. All 18,000 body transforms were checked against native MuJoCo. Target poses are geometric references, not independently simulated outcomes. Root position and orientation are fixed in the schematic; no foot contact, hardware behavior or balance conclusion is inferred.
Policy: Pollen Robotics alpha_walking.onnx, 50 Hz CPU inference. Simulation: MuJoCo, XML position-actuator fallback, not BAM. The upstream model and depicted media retain noncommercial/share-alike terms; see the media notice. Independent implementation inspired by mishig's Microduck Anatomy; no code or assets were copied from that Space.