Scaling Quadruped World Models

Ramakrishna Kompella · Gödel Machines

These 25 FP32 checkpoints forecast the Unitree Go2’s next 30 body and joint readings at 50 Hz. The root weights require future terrain scans. For recorded encoder and IMU histories with proposed joint targets, download the adapted checkpoint in proprio/. In both cases the output is an open-loop motion forecast.

Files Inputs Checkpoints
Root and sweep/w*/b*/ Measured history, joint targets and encoded scans 20
proprio/ and proprio/w*/ Measured history and joint targets 5

The root checkpoint has 120,370,206 per-step parameters. The adapted flagship has 119,670,302 parameters in total. Smaller parents have widths 1,024, 1,440, 2,048 and 2,880. Each directory has a config and a checkpoint. The scan family can use the original PyTorch training checkpoint (weights, training arguments, update count and evaluation scores) or an equivalent SafeTensors export. Adapted models default to SafeTensors. The publisher can also reuse a verified original PyTorch file for an adapted model. The published manifest identifies the file to load. The manifest gives paths, tensor-file hashes, source hashes, exposure and the parent of each adapted model. The 20 parent checkpoints share five training trajectories.

Interface

Use 51 readings including the forecast origin, with the 50 joint targets that connect them. A sequence of 150 future targets returns 150 readings over three seconds. Each reading contains world up in the body frame, three body angular rates (rad/s), twelve joint angles (rad) and twelve joint velocities (rad/s). Joint targets are absolute angles, ordered by hip, thigh and calf, with front left, front right, rear left and rear right within each group. The scanner supplies a 17 × 11 yaw-aligned, clipped terrain encoding at each future recorded pose. No scan is required by the adapted models. Position, linear speed and foot contacts are outside the output interface.

Evaluation

On the scan-conditioned simulation sweep the root checkpoint scores 0.364 three-second normalized endpoint error. On 32 separately held-out flat-ground worlds the adapted flagship scores 0.613 at one second and 1.009 at three seconds. Across nine default-mode real Go2 SDK recordings its three-second joint-angle RMSE is 0.0641 rad. On the four 0.5 m/s recordings the corresponding value is 0.0461 rad. The training data are simulated and the physical recordings come from one Go2 on a flat floor. The physical evaluation replays previously published joint targets.

Use the companion project for the loader, worked examples and paper. Project repository

Downloads last month
-
Video Preview
loading