AN EXPERIMENT BY AJAY CYRIL
How far ahead can a
robot world model predict?
I gave a learned dynamics model and two MuJoCo backends the same starting state and motor commands. Then I measured how their predictions separated.
From one step to one second
Median root-position gap to MuJoCo Warp across 20 windows per friction setting.
Learned model vs. WarpNative MuJoCo vs. Warp
All comparisons and seed ranges
| Look-ahead | Learned ↔ Warp | Learned ↔ native | Native ↔ Warp |
|---|
Replay one measured trajectory
Three predictions from the same start, 0.5 seconds into the seed 42 run. The selected foot friction applies here too.
Loading recorded Warp measurements…
These gaps describe simulator agreement. Physical accuracy would require a measured G1 trajectory.
Read the experiment and its limits → · Download dataRun your own browser comparison
Change the scripted motion or foot friction and rerun native MuJoCo against the learned model.
NATIVE MUJOCO WASM × COMMUNITY WORLD MODELReady
POSITION GAP
—cm
LARGEST GAP
—cm
POSITION GAP OVER TIME
Joint-angle difference: —°