Microphone · two songs
From Dai folk to Locking.
Dai folk music, a short silence, then Locking. The dance changes with the music.
AHEADance
Dances to what
it hears.
A humanoid robot that improvises its dance as the music plays, in the music's own style.
A Unitree G1 dancing to eight songs in seven styles, each played from an audio file. One model for all of them; it is never told the style.
How it works
Listen. Improvise. Dance.
One model, listening as the song plays. Nobody tells it the style or the tempo: it hears them in the music and dances them. Every performance is improvised, so the same song becomes a new dance each time.
- 1 · Listen
It hears the style.
AHEADance listens to the music as it plays, from a file or through a microphone, and hears its style, rhythm and energy for itself.
- 2 · Improvise
A new dance every time.
From what it hears, it creates the robot's next moves for all 29 joints, thirty times a second, even for songs it has never heard.
- 3 · Dance
On a real humanoid.
The robot dances while the music plays. When the music stops, it settles back to standing; when the music returns, it dances again.
Real Unitree G1
Music in the room. A robot that listens.
A loudspeaker plays the music and a microphone picks it up. The robot dances to whatever it hears, in the style of that music.
Microphone · unseen music
A song it never trained on.
Swoop, a Popping track from the test set, through the same loudspeaker and microphone.
Real Unitree G1 · eight songs
One model. Many styles.
Eight runs on the real robot, one per song or set of songs. The model is never told the style; the dance follows what it hears.
One song
Miao
miaozufengcai
3 songs in a row
Dai → Locking → Hip-hop
shuiyueliang → Runaway Baby → Peaches
One song
Chinese
juelian
One song
K-pop
LOCO
One song
Urban
suffer
One song
Choreography
tebiederen
3 songs in a row
Hip-hop → Hip-hop → Breaking
GoCrazy → Ikarus → Just Begun
One song
K-pop
UhOH
In these runs the music file streams to the model as it plays, and a loudspeaker plays the same music in the room. Style names are the FineDance music labels.
The same eight runs, lined up at the moment their music stops.
Live
Bring your own music.
Play any song from your device, or play it on your phone and let the page listen through the microphone. Our model on a cloud GPU listens as it plays, and the robot dances to it as on the real robot. Every session is a new dance.
- Any songa music file, or your phone through the microphone (beta)
- 1 minuteper session
- 2 sessionsa day for each email
- About a minuteto start the GPU if no one has used it for a while
Web demo
Play it in your browser.
Pick a dance from the list. The music plays here, and a model of the robot replays what the model generated for exactly that audio, with the same timing as on the real robot. Under Same song, one song is danced three times; under Live, play your own music file and the model dances to it as it plays.
Drag to turn
Take a closer look
Music in. Motion out.
The dance loaded in the web demo, or your live session as it happens, on one timeline: what the robot heard, and how much it moved.
Loaded dance
MusicRobot activity (mean joint speed)
Music
Robot
- One run, exactly this audio.Every dance in the web demo is one run of the deployed model on the audio you hear: seed 1234, except the three Same song takes (seeds 1234, 1235, 1236) and Live (a new seed each session). Nothing is cut together from other runs.
- Tracker or generated motion.By default a replay shows the generated motion followed by the whole-body motion tracker our robot runs, in MuJoCo physics, so the robot's own collisions and balance apply. Generated motion shows what the model output, as it is. Neither is the joint angles measured on the real robot. Live shows the tracker too. If the tracker falls or stops, the page says so and shows the generated motion. The songs you send show the generated motion.
- Robot status.The label on the stage comes from the joint speed of that motion: below 0.12 rad/s is standing, above 0.25 rad/s is dancing.
- Styles are only labels.Style names are the FineDance music labels. The model only hears the music.
Your song
Send us 30 seconds.
We put 2 s of silence before your music and 4 s after it, so the robot starts and ends standing. Then we run the same model on it on a cloud GPU, and the dance appears in your web demo under Your song. Only you and our lab see your songs.
Sign in with your email to send songs and to go live.
·
Each email can send 2 songs and go live twice a day. The code is sent from AHEADance <aheadance@gmail.com>; please also check your junk email. We use your email only to send the code and keep your songs; only our lab sees it.