Supplementary results / G1 locomotion
Seven native terrain families
The same three saved policies are used in every row: one training seed per actor, five prescribed geometry/reset groups and 100attempts per cell. All 21cells / 2,100 episodes come from the fully validated broader sweep. No checkpoint is selected separately by terrain or actor result.
Nominal evaluation physics, 0.5 m/s forward command and 20 seconds from rest. Survival excludes native orientation falls and leaving the certified support footprint. RMSE and forward distance include all attempted episodes, including early failures. Compound success additionally requires linear and yaw RMSE≤0.2 anddirected progress≥80% of commanded displacement.
| Terrain family | Actor | 20s survival | Compound success | Linear RMSE m/s | Yaw RMSE rad/s | Forward m | Support exits | Native falls |
|---|---|---|---|---|---|---|---|---|
| flat | PRISM | 100/100 | 0/100 | 0.1115 | 0.2843 | 11.021 | 0 | 0 |
| pyramid_stairs | PRISM | 100/100 | 0/100 | 0.1137 | 0.3362 | 7.404 | 0 | 0 |
| pyramid_stairs_inv | PRISM | 100/100 | 0/100 | 0.1229 | 0.3609 | 8.892 | 0 | 0 |
| hf_pyramid_slope | PRISM | 77/100 | 0/100 | 0.2324 | 0.5855 | 3.679 | 0 | 23 |
| hf_pyramid_slope_inv | PRISM | 99/100 | 0/100 | 0.1274 | 0.7908 | 0.948 | 0 | 1 |
| random_rough | PRISM | 98/100 | 0/100 | 0.1606 | 0.6043 | 6.937 | 0 | 2 |
| wave_terrain | PRISM | 100/100 | 0/100 | 0.1240 | 0.3973 | 10.402 | 0 | 0 |
| flat | MLP | 100/100 | 0/100 | 0.1294 | 0.3776 | -1.292 | 0 | 0 |
| pyramid_stairs | MLP | 100/100 | 0/100 | 0.1314 | 0.4241 | 4.108 | 0 | 0 |
| pyramid_stairs_inv | MLP | 100/100 | 0/100 | 0.1226 | 0.4961 | -0.086 | 0 | 0 |
| hf_pyramid_slope | MLP | 97/100 | 0/100 | 0.3320 | 0.4078 | 2.765 | 0 | 3 |
| hf_pyramid_slope_inv | MLP | 96/100 | 0/100 | 0.1586 | 0.7007 | 0.805 | 0 | 4 |
| random_rough | MLP | 82/100 | 0/100 | 0.1825 | 0.6606 | 3.543 | 0 | 18 |
| wave_terrain | MLP | 99/100 | 0/100 | 0.1182 | 0.4824 | 6.008 | 0 | 1 |
| flat | Matched | 100/100 | 0/100 | 0.1016 | 0.4573 | -2.409 | 0 | 0 |
| pyramid_stairs | Matched | 100/100 | 0/100 | 0.1143 | 0.4778 | 0.804 | 0 | 0 |
| pyramid_stairs_inv | Matched | 83/100 | 0/100 | 0.1413 | 0.5185 | -0.638 | 2 | 15 |
| hf_pyramid_slope | Matched | 100/100 | 0/100 | 0.1664 | 0.4884 | 3.959 | 0 | 0 |
| hf_pyramid_slope_inv | Matched | 100/100 | 0/100 | 0.1198 | 0.8547 | 1.159 | 0 | 0 |
| random_rough | Matched | 61/100 | 0/100 | 0.2391 | 0.8026 | 0.133 | 7 | 32 |
| wave_terrain | Matched | 100/100 | 0/100 | 0.1201 | 0.4721 | 8.301 | 0 | 0 |
The focused random-rough comparison was chosen after inspecting this seven-terrain sweep. The pyramid profile starts on a central platform and can enter a descending section first; it is not a continuous uphill test. PRISM had 23 early orientation falls in the first descending section of hf_pyramid_slope; survival there was 77/100, versus MLP 97/100 and Matched 100/100.
Training configuration
All methods use the same native-mjlab task and 30,000-update budget. Startup samples combine 20% nominal physics with 80% independent friction 0.3–1.2, link mass/inertia ratio 0.8–1.2 and fixed torso payload 0–5 kg. Shared final-hidden RMSNorm is applied equally to all three actors.