A humanoid that dances to streamed music must commit each motion chunk before it has heard the music that chunk accompanies, and must continue from motion it has already committed. We present AHEADance, a robot-native streaming dance generator executed on a humanoid robot by a whole-body tracker. ListenAhead conditions each chunk on heard music and a frozen music model's forecasts of the coming music, aligned to their target times and weighted by reliability. The forecasts carry the timing of upcoming beats almost as well as the actual future; with them, the generator is better synchronized with the music and concentrates more motion energy at its song's beat frequency. Commit Forcing trains continuation from the model's own committed chunks, updating both the latent history and the decoded boundary of the next chunk; it reduces low activity from 83.5% to 10.3%, the level of real dance. Three music evaluators, each validated on real G1 dance, test synchronization, song and style. On a physical Unitree G1, it dances while the music plays and comes to rest when it stops, without a fall. Project page: https://ihcr.github.io/AHEADance
Figures & tables
Fig. 2: Overview of AHEADance. ListenAhead builds a time-aligned music condition from heard audio and reliability-weighted forecast states. Each chunk is generated in three stages (root trajectory, discrete codes, continuous corrections), decoded once by the commit-aligned motion codec from the committed boundary, committed, and tracked by a whole-body tracker. Commit Forcing trains continuation from the model’s own committed chunks. Robot renderings are schematic, except in the Physical panel.
Method
Audio
Causal
Robot ref.
Own hist.
Forecast
EDGE [ 3 ]
✓
—
—
—
—
LODGE [ 6 ]
✓
—
—
—
—
DiscoForcing [ 7 ]
✓
✓
—
—
—
InfiniDance [ 9 ]
✓
✓
✓
— †
—
RoboPerform [ 8 ]
∗
✓
—
—
—
AHEADance
✓
✓
✓
✓
✓
TABLE I: Interfaces of Music-Driven Methods
Fig. 3: Commit-consistent continuation training. Reference-history training supplies recorded contexts. Commit Forcing reconstructs a sequence of K perturbed reference chunks, decodes and commits each complete chunk, and carries the updated context into the next reconstruction. Each update pairs generated history with the boundary state of that same decoded motion. This history construction stops gradients; the current continuation is supervised with reference targets. The inset uses the shared code-relative target rule of Eq. 4 .
Method
FID k↓
Div k
Sync ↑
MMR ↑
Tempo ↑
Real dance
–
6.22
69.0
0.245
80.2
Online (no future music)
AHEADance
18.8
2.22
28.6
0.187
57.4
Bailando++, strict
65.2
0.59
11.9
− 0.010
48.1
Danceba, strict
44.8
4.52
9.5
− 0.003
46.6
MEGADance, strict
23.7
3.77
11.9
− 0.011
49.4
TABLE II: Comparison on 42 Unseen Songs
Music
Continuation
Motion distribution
Setting
ChoreoSync ↑
MMR ↑
BAS ↑
Tempo ↑
Low act. ↓
Span ↑
FID k↓
FID g↓
Div k
Real dance
45.2
0.184
0.247
71.3
9.7
0.922
–
–
6.28
AHEADance
24.2
0.126
0.262
56.8
10.3
0.732
16.7
46.3
2.94
w/o future channel
17.4
0.054
0.268
51.3
17.5
0.604
21.3
74.1
3.11
w/o forecasts (last heard state)
15.8
0.047
0.265
48.3
10.8
0.636
23.2
71.1
2.55
Teacher Forcing
18.4
0.023
0.238
47.6
83.5
0.492
64.6
140.6
1.53
TABLE III: Training Ablations on 721 Held-Out Sequences
Score
AIOZ
FineDance
AIST++
CoMPAS3D
All
Chance
Sequences
506
103
95
17
721
ChoreoSync (%)
46.2
26.2
53.7
82.4
45.2
11.1
MMR (%)
68.4
68.7
67.6
80.9
68.6
50
Tempo (%)
71.2
68.1
79.3
45.9
71.3
50
Style (%)
–
84.2
65.3
–
–
33 / 10
BAS, own music
0.247
0.216
0.283
0.212
0.247
–
TABLE IV: Score Validation on Real G1 Dance
Fig. 4: ListenAhead. (a) What the forecasts carry: cross-validated linear read-outs (Pearson r on held-out songs; five song folds, 235 songs) of music descriptors, by how far ahead, from MRT-2 states of the heard music, the heard music with the forecasts, or the heard music with the actual future (a non-causal upper bound). (b) What the motion uses: tempo salience of three generators (without the future channel, with it holding the last heard music state, and with the forecasts) on AIST++ music (60 pieces), ten of these pieces time-stretched to five speeds between 0.8 × and 1.2 × , and songs unseen in training (AIOZ-GDANCE 42, FineDance 30). ∗ , † : the 95% CI of the paired difference over songs from the generator without the future channel ( ∗ ) or with the last heard state ( † ) excludes zero (t interval for the ten time-stretched pieces). Diamonds: real dance of the same songs (51 AIST++ pieces and 27 FineDance songs; none for time-stretched music); dashed line: chance.
Fig. 5: One chunk on the robot, relative to its audio cutoff τ : medians over the 12 physical runs. Dashed: commit deadline (1.24 s delay minus the 0.9 s preview); open circle: commit time if each chunk’s music were heard before generating it.
Fig. 6: The physical Unitree G1 dancing to songs of seven styles and standing after the music stops (backgrounds lightened).
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Setting
ChoreoSync ↑
MMR ↑
BAS ↑
Tempo ↑
Low act. ↓
Span ↑
FID k↓
FID g↓
Div k
AHEADance (D+C)
24.2
0.126
0.262
56.8
10.3
0.732
16.7
46.3
2.94
Discrete only (D)
20.4
0.078
0.283
51.4
15.1
0.676
26.2
64.9
2.44
Continuous only (C)
28.0
0.054
0.270
57.4
34.6
0.589
28.5
81.3
2.60
Appendix
TABLE V: Chunk Representation on the Same 721 Sequences
Real dance
721 sequences
Quality
Split
1 s loop
AHEADance
w/o future
TF
w/o closure
D
C
FID LMA
15.2
1443.7
33.6
41.7 +
103.3 +
76.7 +
47.1 +
52.3 +
Appendix
TABLE VI: FID LMA and Laban Movement Qualities
Method
Trained on
FID k↓
Div k
Sync ↑
MMR ↑
BAS ↑
Tempo ↑
Real dance
–
–
6.22
69.0
0.245
0.247
80.2
Online (no future music)
AHEADance
6 datasets
18.8
2.22
28.6
0.187
0.266
57.4
Bailando++, strict
AIST++
153.5
14.77
11.9
0.002
0.250
56.6
Danceba, strict
AIST++
134.8
16.51
11.9
− 0.006
0.261
50.9
MEGADance, strict
FineDance
79.1
8.44
11.1
0.003
0.251
48.2
Appendix
TABLE VII: Released Weights on the Same 42 Songs
Method
Trained on
Music
Acc.
Follow ↑
AHEADance − X
Real dance
–
–
100.0
47.4
–
Online (no future music)
AHEADance
6 datasets
online
85.5
32.6
–
Bailando++, strict
AIST++
online
39.1
3.0
29.6[10.1,49.0]
Danceba, strict
AIST++
online
60.9
6.7
25.9[8.7,45.5]
MEGADance, strict s
FineDance
online
60.9
− 0.2
32.8[12.8,54.5]
Appendix
TABLE VIII: Style on 23 FineDance Songs Unheard by AHEADance
Robot motion generation creates feasible, task-directed trajectories for robots to move, manipulate objects, and interact with people and environments.
School of Artificial Intelligence, Beijing University of Posts and Telecommunications · Beijing Academy of Artificial Intelligence · School of Software and Microelectronics, Peking University