⟨GÖDEL MACHINES⟩

Scaling Quadruped World Models

A 120-Million-Parameter Predictive Model of the Unitree Go2

Ramakrishna Kompella

GÖDELMACHINES
One walk, three views. The Go2 received a 0.5 m/s command toward a 2 m odometry target. The 120M model and MuJoCo replay its recorded joint targets for 5.08 s. The plot tracks the front-left thigh against its encoder. Across all 12 joints, RMSE was 0.0444 rad for the neural forecast and 0.0628 rad for MuJoCo with the recorded gains.

Both rendered bodies are shown at a fixed position and heading to compare joint motion and tilt. The neural model has no global position output. The iPhone view is aligned to command onset using host timestamps. Camera transport delay is unknown. Full camera recording, including settling

3D pose overlay

Rotate the three poses together. The dark skeleton comes from the Go2’s joint encoders and IMU. Orange and teal show where the 120M model and MuJoCo place the same joints and feet under the recorded commands.

Loading…

Watch overlay as MP4 · Download 3D coordinates

3D foot-center error against the telemetry reconstruction (cm RMS)
Recorded trial120M neuralMuJoCo
2 m walk · 5.08 s2.763.65
StaticWalk · 1.30 s6.2018.27
TrotRun · 2.00 s8.0944.15

Each row is one recording. We pool the squared 3D distances for all four foot centers across forecast frames, then take the square root. The initial pose is shared and excluded.

All layers use the robot’s measured heading, controller odometry and reported body height. Each layer supplies its own tilt and 12 joint angles. Foot positions come from forward kinematics at 50 Hz, so their separation measures limb and tilt error under that shared placement. The traces follow the selected foot center over the preceding 0.4 s. The floor grid is spaced at 2 ft. Unitree Go2 URDF

Abstract

How do model size and training exposure affect a quadruped’s ability to forecast motion from joint commands? Five recurrent predictors of the Unitree Go2, ranging from 8M to 120M parameters, forecast 30 joint and body measurements at 50 Hz. Across four training budgets, one scaling law describes three-second prediction error when recorded terrain scans are available. Withholding the final exposure doubling gives a mean relative prediction discrepancy of 2.9%.

Removing the scan produces a model that can run from joint targets and proprioceptive history alone. Updating its input projection improves the 120M model’s one-second error on held-out flat ground from 0.831 to 0.613. At three seconds the error changes from 0.948 to 1.009. On nine physical Go2 trials, the 8M model gives the lowest pooled joint-angle error among the five adapted sizes. On the four walks commanded at 0.5 m/s, the 120M model gives the lowest error.

Predicting the next doubling

Five model sizes and four training budgets yield 20 checkpoints. With the largest budget at each width withheld, the fitted law predicts three-second endpoint error with 2.9% mean relative discrepancy. Carrying the preceding scores forward gives 7.4%.

L3s = 0.202 + 0.00775 n−0.863 + 0.229 d−0.246

Here n is size in units of 30M per-step parameters and d is exposure in billions of sampled transitions. The fit uses recorded terrain scans in simulation.

Measured three-second error by exposure and estimated training compute for all five widths
24,576 evaluation windows across 96 held-out shards. The unstable 8M branch is included.

Replaying the robot’s commands

The stock controller runs on the real robot once. Each predictor receives its published joint targets, and the encoders and IMU supply the reference. Across nine default-mode runs, the smallest adapted model has the lowest pooled joint-angle error. The ordering changes on the four faster walks.

Three-second joint-angle RMSE (rad)
PredictorNine SDK runsFour at 0.5 m/s
8M adapted0.05660.0512
120M adapted0.06410.0461
MuJoCo, recorded gains0.05860.0729
Paired joint-angle errors by predictor for the nine individual SDK recordings and pooled result
Each gray line joins predictions for one physical trial.

On the filmed walk above, the stationary foot-contact centroid moved an estimated 2.335 m on the two-foot grid. Settled onboard odometry was 2.035 m. The measurements refer to different points on the robot.

StaticWalk request. The replay covers 1.30 s. The controller reached its distance stop at 0.459 m in 1.31 s. Across the 12 joints, RMSE was 0.128 rad for the neural model and 0.276 rad for MuJoCo. Full camera recording
TrotRun request. The replay covers 2.00 s. At the 2.01 s time limit, controller odometry read 0.362 m. MuJoCo rolls as the real robot stays upright. Mean body-up error was 77.7° for MuJoCo and 7.4° for the neural model. Full camera recording

Each clip shows one acknowledged SDK mode request. The 120M model receives the published joint targets. MuJoCo also receives the recorded gains, velocity targets, and feedforward torques. Both replays show pose in place. The gait labels come from the SDK requests. Independent footfall classifications are unavailable.

Eight gaits in simulation

MuJoCo and the T2 policy supply a three-second reference for each gait. The adapted 120M model receives the joint targets from that same trajectory. Trot gives a mean body-up error of 0.52°. In pronk, the simulator overturns and the error reaches 69.52°.

See the per-gait measurements →

Joint, gyro and up-direction errors through three seconds for eight simulated gaits
Instantaneous sensor errors in simulation. Each row uses a common logarithmic scale across its two panels.

BibTeX

@misc{kompella_scaling_go2_2026,
  title = {Scaling Quadruped World Models: A 120-Million-Parameter Predictive Model of the Unitree Go2},
  author = {Kompella, Ramakrishna},
  year = {2026},
  institution = {Gödel Machines}
}

Download citation