TL;DR
Quantitative comparison of 4 baselines on
17,348 clips — TMR semantic retrieval
(R@K, MM-Dist, CLIP, FID, diversity) and physical plausibility
(accel, jerk, penetration). Larger radar polygon = better;
lower-is-better axes are inverted so all axes share the same direction.
⚠ The aggregate radars and tables collapse very different action regimes into one number and obscure where each method actually wins or fails. Use the Per-domain breakdown at the bottom of the page to see how the ranking changes across domains (fine manipulation, locomotion, sports, etc.) — that's where the interesting story lives.
⚠ The aggregate radars and tables collapse very different action regimes into one number and obscure where each method actually wins or fails. Use the Per-domain breakdown at the bottom of the page to see how the ranking changes across domains (fine manipulation, locomotion, sports, etc.) — that's where the interesting story lives.
Aggregate semantic radar
How to read: each axis is mapped to a fixed
[bad, ideal] band tuned per-metric, so polygon
size reflects absolute quality, not cohort rank.
Lower-is-better axes (FID, MM-Dist) are inverted —
larger polygon is always better. The grey
Ground Truth polygon is the upper-bound reference;
read each baseline as the gap to that envelope.
Hover any vertex for raw values;
hover this caption for the exact band per axis.
Aggregate semantic metrics — TMR retrieval
| Method | R@1 ↑ | R@3 ↑ | R@10 ↑ | MM-Dist ↓ | CLIP ↑ | FID ↓ | Div (gen) ↑ |
|---|---|---|---|---|---|---|---|
| Ground Truth | 61.8% | 81.6% | 95.2% | 71.014 | 0.578 | 0.000 | 120.5 |
| MDM | 47.2% | 71.3% | 91.8% | 77.946 | 0.484 | 0.062 | 116.7 |
| MoMask | 51.7% | 76.7% | 94.1% | 76.242 | 0.507 | 0.105 | 115.7 |
| Kimodo | 38.9% | 62.8% | 87.5% | 81.271 | 0.425 | 0.106 | 114.6 |
| MotionMillion | 29.4% | 51.7% | 80.3% | 87.879 | 0.349 | 0.276 | 109.2 |
italic gray rows — ground-truth / external MoCap reference (not ranked)
bold green — best in column among baselines
arrows: ↑ higher is better, ↓ lower is better
Physical plausibility
How to read: every axis is lower-is-better with the
radar rim anchored at clean external MoCap, not at RoMo
GT — so smoother-than-GT methods score above the GT polygon
instead of being penalised for it.
Hover this caption for the exact bounds and reference policy.
| Method | Accel (m/s²) ↓ | Intra-Jerk (m/s³) ↓ | Inter-Pen ↓ | Self-Pen MAX (mm) ↓ | HF Power ↓ |
|---|---|---|---|---|---|
| Ground Truth | 8.379 | 14.259 | 0.64% | 0.002 | 0.0081 |
| GRAB GT (ref.) | 3.362 | 15.920 | — | — | 0.0011 |
| GigaHands GT (ref.) | 4.963 | 35.418 | — | — | 0.0132 |
| MDM | 7.915 | 71.176 | 2.19% | 0.382 | 0.0132 |
| MoMask | 4.992 | 265.218 | 1.58% | 0.017 | 0.0153 |
| Kimodo | 6.032 | 22.893 | 1.06% | 0.066 | 0.0077 |
| MotionMillion | 5.766 | 76.097 | 1.82% | 0.000 | 0.0082 |
italic gray rows — ground-truth / external MoCap reference (RoMo GT, GRAB, GigaHands)
bold green — best in column among baselines
arrows: ↑ higher is better, ↓ lower is better
Per-domain breakdown
This is where the aggregate hides the interesting story.
Pick one or more domains below to see the per-category
ranking — method order can flip across fine manipulation,
locomotion, sports, etc. Tighter
[bad, ideal] bands
are used here so methods don't collapse to the chart centre on
the smaller per-domain pools.
Hover this caption for the exact bounds and physical-grouping policy.
Metric set
R@1 / R@3 / R@10 / MM-Dist / CLIP / FID
Semantic
6 axes
Physical
5 axes
Domain