Multi-step Kinematic and Semantic Rollout on AIDE (test)
52.89MPJPE (All, px)Zero-Velocity
Evaluation results
| Method | Links | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Zero-VelocityRollout Protocol=5→5, Seeds=42, 123, 456, Baseline Type=Motion-only & offline, Note=copy-last-frame baseline2026.05 | 52.89 | 2.4 | 139.19 | 6.32 | 40.72 | 8.27 | 59.61 | 14.94 | — | — | — | — | 85.95 | 93.61 | |
| KL bottleneckRollout Protocol=5→5, Seeds=42, 123, 456, Variant=Driver-WM variant2026.05 | 71.53 | 3.25 | 137.85 | 6.26 | 73.89 | 67.43 | 78.38 | 71.08 | 92.67 | 89.13 | 82.27 | 66.4 | 71.17 | 88.31 | |
| Driver-WM (main)Rollout Protocol=5→5, Seeds=42, 123, 4562026.05 | 71.66 | 3.25 | 138.34 | 6.28 | 73.34 | 69.09 | 80.35 | 73.08 | 92.94 | 89.26 | 82.81 | 66.54 | 71.4 | 88.21 | |
| w/o ext contextRollout Protocol=5→5, Seeds=42, 123, 456, Variant=Driver-WM variant2026.05 | 71.74 | 3.26 | 136.94 | 6.21 | 73.4 | 71.33 | 74.22 | 63.56 | 65.52 | 26.39 | 53.86 | 14 | 71.13 | 88.2 | |
| +Skeleton-PreRollout Protocol=5→5, Seeds=42, 123, 456, Variant=Driver-WM variant2026.05 | 72.32 | 3.28 | 138.35 | 6.28 | 72.58 | 68.04 | 80.51 | 73.98 | 92.77 | 89.01 | 81.88 | 69.57 | 70.93 | 87.94 | |
| MotionBERTRollout Protocol=5→5, Seeds=42, 123, 456, Baseline Type=Motion-only & offline2026.05 | 73.74 | 3.35 | 142.61 | 6.47 | 66.77 | 54.96 | 66.28 | 52.37 | — | — | — | — | 77.77 | 92.89 | |
| RSSM/GRU dyn.Rollout Protocol=5→5, Seeds=42, 123, 456, Variant=Driver-WM variant2026.05 | 76.48 | 3.47 | 137.21 | 6.23 | 71.53 | 67.7 | 79.86 | 72.29 | 92.34 | 87.74 | 81.5 | 63.91 | 68.46 | 86.97 | |
| Offline Enc-DecRollout Protocol=5→5, Seeds=42, 123, 456, Baseline Type=Motion-only & offline2026.05 | 76.91 | 3.49 | 145.14 | 6.59 | 66.34 | 50.61 | 64.2 | 50.56 | — | — | — | — | 76.78 | 92.69 | |
| Cross-Attn onlyRollout Protocol=5→5, Seeds=42, 123, 456, Interface=VLM interface2026.05 | 79.27 | 3.6 | 141.23 | 6.41 | 66.61 | 61.31 | 77.45 | 69.27 | 92.12 | 88.18 | 83.36 | 68.72 | 66.75 | 85.85 | |
| Late fusionRollout Protocol=5→5, Seeds=42, 123, 456, Interface=VLM interface2026.05 | 79.93 | 3.63 | 140.89 | 6.39 | 68.42 | 53.02 | 72.25 | 54.23 | 92.45 | 88.23 | 82.92 | 67.99 | 65.98 | 86.04 | |
| Single-streamRollout Protocol=5→5, Seeds=42, 123, 456, Interface=VLM interface2026.05 | 92.26 | 4.19 | 146.98 | 6.67 | 45.32 | 13.7 | 60.43 | 18.08 | 65.52 | 26.39 | 53.86 | 14 | 59.19 | 81.57 |