Dataset card

README.md

@@ -0,0 +1,37 @@
+---
+license: cc-by-4.0
+task_categories:
+- robotics
+tags:
+- robotics
+- music
+- imitation-learning
+- mujoco
+- unitree-g1
+pretty_name: Kraftwerk Bot demonstrations
+viewer: false
+---
+
+# Kraftwerk Bot: drumming takes and robot demonstrations
+
+The data for [github.com/diw22/DWinters_Humanoid](https://github.com/diw22/DWinters_Humanoid): a simulated Unitree G1
+humanoid that hears a drum loop with one part missing and plays that part on an Akai LPD8 (claps on pad 1, rims on
+pad 5, a shaker fade on knob 4).
+
+Download it into the code's `data/` folder:
+
+```bash
+hf download diw22/DWinters_Humanoid-data --repo-type dataset --local-dir data
+```
+
+| Folder | Contents |
+|---|---|
+| `human/<take>/` | 160 takes of me playing: `human.mp4` (+ `human_cam2.mp4`, a second angle), `played.wav` (my part), `music.wav` (what the robot hears: every part but its own, 24 kHz), `episode.json` (part, split, hit times, camera calibration), the cached hand tracking; shakers takes also have `part.wav` and `knob_target.npz` |
+| `robot/<take>/` | 158 simulated G1 demonstrations made from those videos: `robot.mp4` (the robot's head camera, lossless), `robot.npz` (joint states, actions, knob angles), `robot.json` (hits, checks, controller layout), `compare.mp4` (me / robot, with sound) |
+| `music/` | the song's loops (kick, claps, rims, shakers, 6 chord loops; 44.1 kHz stereo); training mixes each take's chord variations from them |
+| `jaco_v1/<take>/` | 49 demonstrations from version 1 (claps on two simulated Kinova Jaco arms) |
+
+152 robot demonstrations pass the physics check and are used for training (split train / validation / test: claps
+34 / 7 / 7, rims 40 / 9 / 8, shakers 34 / 7 / 6). The demonstrations were simulated with MuJoCo 3.3.0.
+
+Recordings and music by Dylan Winters, licensed CC BY 4.0.