Robot policies
Every policy here was trained in RL Studio from a description or a reference video, and exported as ONNX with its observation normalizer baked in — download one and run it on the robot. Hover a card to watch the policy.
Imitate video — Microduck
Walk forward at a steady pace
Microduck·anonymous0 downloads
G1 — first strides
Velocity-tracking walk on the corrected G1 profile: upright alternating strides with arm swing, trained 4000 iterations.
Unitree G1·Luigi D'Introno26 downloads
Microduck — wander
Waddles around in lazy curves, changing direction as it pleases.
Microduck·Priya N.7 downloads
Microduck — happy march
Marches in place with a bobbing head — trained on the champion gait clock with go-nowhere commands.
Microduck·Sam Ortiz9 downloads
Microduck — turn on command
Rotates in place following yaw commands — steady continuous turning with quick little steps.
Microduck·Maya Chen15 downloads
Open Duck — bounce
Rhythmic crouch-and-rise bounce learned from a jumping reference video.
Open Duck Mini v2·Robin K.12 downloads
Open Duck — turn on command
Rotates in place following yaw commands, repositioning with small quick steps.
Open Duck Mini v2·Jonas W.17 downloads
Open Duck — walk
Steady forward walk trained on velocity commands — upright, alternating steps, keeps its line.
Open Duck Mini v2·Priya N.21 downloads
Microduck waddle — closed-loop tuned
The first policy that actually walks: 8 measured diagnose-and-adjust cycles on a phase-locked gait recipe. 0.5 falls/min, real anti-phase stepping.
Microduck·Luigi D'Introno34 downloads