Datasets:

joonsu0109
/

soundact_data

---
license: cc-by-4.0
pretty_name: Egocentric stereo-audio robot manipulation demonstrations (T1-T4)
task_categories:
- robotics
tags:
- robotics
- audio-visual
- stereo-audio
- sound-source-localization
- imitation-learning
- manipulation
- real-robot
- zarr
---
# Egocentric stereo-audio robot manipulation demonstrations (T1–T4)

Real-robot teleoperated demonstrations in which a robot arm uses a stereo microphone pair and an RGB camera on its wrist to find a sound source that starts out of the camera's view and then manipulates it.

- 4 tasks, 311 episodes, 136,954 steps (about 114 minutes) at 20 Hz.
- Per step: wrist RGB (policy input), wrist depth, external RGB and depth, 50 ms of stereo audio at 48 kHz, end-effector pose and gripper width.
- Robot: Franka Research 3, teleoperated with a gamepad. Cameras: Intel RealSense RGB-D.

## Files

| Path in this repository | Task | Original dataset name | Size |
|---|---|---|---|
| `T1_ring_off/replay_buffer.zarr.zip` | T1 Ring-off | `ral_v3_timer_ring_off_out_sight` | 1.50 GB |
| `T2_covered_phone/replay_buffer.zarr.zip` | T2 Covered phone | `rss_coverd_iphone` | 2.74 GB |
| `T3_call_bell/replay_buffer.zarr.zip` | T3 Call bell | `rss_short_time_bell` | 2.44 GB |
| `T4_animal_feeding/replay_buffer.zarr.zip` | T4 Animal feeding | `rss_dog_cat_duck` | 3.02 GB |

Sizes are in decimal GB (10^9 bytes).

## Tasks

- **T1 Ring-off.** A kitchen timer rings somewhere on a semicircular workspace around the robot, out of the wrist camera's view at the start. The robot turns toward the sound and presses the timer to stop it.
- **T2 Covered phone.** A phone rings under one of several identical towels. The robot finds the towel the sound comes from and pulls it off to uncover the phone.
- **T3 Call bell.** A call bell rings briefly. The robot goes to the bell, picks up the fork lying next to it and puts the fork into the cutlery tray.
- **T4 Animal feeding.** A speaker next to a bowl plays a dog, cat or duck call. The robot picks food from three trays at fixed positions (the animal call decides which tray) and drops it into the bowl (the direction of the call decides where).

| Task | Original name | Episodes | Steps | Episode length, mean (min–max) | Left / Front / Right | Held-out episodes (split seed) | Recorded |
|---|---|---|---|---|---|---|---|
| T1 | `ral_v3_timer_ring_off_out_sight` | 86 | 24,332 | 14.15 s (10.25–17.30 s) | 43 / 3 / 40 | 6, 7, 17, 35, 52, 59, 61, 71, 81 (seed 42) | 2026-09-09, 22:59–23:48 KST |
| T2 | `rss_coverd_iphone` | 80 | 36,431 | 22.77 s (19.40–27.60 s) | 39 / 3 / 38 | 3, 16, 18, 26, 51, 52, 62, 67 (seed 15) | 2026-10-06, 20:37–22:14 KST |
| T3 | `rss_short_time_bell` | 70 | 36,597 | 26.14 s (21.20–34.75 s) | 32 / 4 / 34 | 5, 6, 29, 43, 50, 59, 67 (seed 42) | on or before 2026-10-06 (archive written 17:44 KST) |
| T4 | `rss_dog_cat_duck` | 75 | 39,594 | 26.40 s (20.75–35.45 s) | 36 / 3 / 36 | 6, 8, 18, 29, 45, 47, 57, 72 (seed 44) | 2026-10-07, 15:42–17:56 KST |

T4 class blocks (episodes are stored in recording order, 25 per class):

| T4 class | Episodes | Left / Front / Right |
|---|---|---|
| dog | 0–24 | 12 / 1 / 12 |
| cat | 25–49 | 12 / 1 / 12 |
| duck | 50–74 | 12 / 1 / 12 |

Notes on the table:

- **Left / Front / Right** is the azimuth of the target point seen from the end-effector position at the first step, in the robot base frame (+x forward, +y left). |azimuth| ≤ 15° counts as front. The target point is the final end-effector position (T1, press), the first gripper close (T2 towel, T3 fork) or the first gripper re-open (T4, drop into the bowl).
- **Held-out episodes** are the episodes we held out for offline evaluation. They come from the training code's split function with `val_ratio = 0.1`:
  ```python
  import numpy as np
  n_val = min(max(1, round(n_episodes * 0.1)), n_episodes - 1)
  held_out = np.random.default_rng(seed=seed).choice(n_episodes, size=n_val, replace=False)
  ```
- **Recorded** comes from the per-step wall-clock timestamps of the original recordings. These timestamps are not stored in the T3 archive, so only the date the archive was written is known for T3.

## Format

Each file is a zipped zarr (v2) group in the [Diffusion Policy](https://github.com/real-stanford/diffusion_policy) / [UMI](https://github.com/real-stanford/universal_manipulation_interface) `ReplayBuffer` layout. All arrays under `data/` share the first axis: the steps of all episodes, concatenated in order at 20 Hz. `meta/episode_ends[i]` is the exclusive end index of episode `i`.

| Key | Shape | Dtype | Content |
|---|---|---|---|
| `data/camera0_rgb` | (N, 224, 224, 3) | uint8 | Wrist RGB camera. This is the policy's image input. |
| `data/camera0_depth` | (N, 224, 224, 3) | uint8 | Wrist depth as a colorized image (see below). |
| `data/camera1_rgb` | (N, 224, 224, 3) | uint8 | External (third-person) RGB camera. Not a policy input. |
| `data/camera1_depth` | (N, 224, 224, 3) | uint8 | External depth as a colorized image. |
| `data/mic_0` | (N, 2400) | float64 | Left microphone waveform, 48 kHz, 2400 samples (50 ms) per step. |
| `data/mic_1` | (N, 2400) | float64 | Right microphone waveform, same layout. |
| `data/robot0_eef_pos` | (N, 3) | float32 | End-effector position (m) in the robot base frame. |
| `data/robot0_eef_rot_axis_angle` | (N, 3) | float32 | End-effector orientation as an axis-angle vector (rad). |
| `data/robot0_gripper_width` | (N, 1) | float32 | Gripper opening width (m). |
| `meta/episode_ends` | (E,) | int64 | Exclusive end index of each episode. |

- Images are stored one frame per chunk with JPEG-XL compression (`imagecodecs_jpegxl`), so reading them needs `imagecodecs` (see below). Other arrays use Blosc/LZ4.
- Original streams were RGB and depth at 640×480, 30 fps per camera, stereo audio at 48 kHz, end-effector pose at 100 Hz and gripper width at 30 Hz. They were synchronized to 20 Hz, and the images were resized to 224×224.
- The depth arrays are not metric depth. The recording code clips depth to 0.01–2.0 m, scales it linearly to 0–255 and applies the OpenCV VIRIDIS colormap. Pixels with no depth reading map to the colormap value of 0.
- Audio: to get a continuous waveform, flatten the rows of an episode (`mic_0[s:e].reshape(-1)`). The microphone spacing used in our training configuration is 0.12 m (a configuration value, not a measurement).
- There is no separate `action` array or timestamp array. Our training code builds the action from the proprioceptive arrays as `[eef_pos, eef_rot_axis_angle, gripper_width]` at the following steps, as in UMI.

## Loading

```python
# pip install "zarr<3" imagecodecs huggingface_hub
import zarr
from huggingface_hub import hf_hub_download
from imagecodecs.numcodecs import register_codecs
register_codecs()  # registers the JPEG-XL codec used by the camera arrays
path = hf_hub_download(
    repo_id="joonsu0109/soundact_data",
    repo_type="dataset",
    filename="T1_ring_off/replay_buffer.zarr.zip",
)
store = zarr.ZipStore(path, mode="r")
root = zarr.open(store, mode="r")
ends = root["meta/episode_ends"][:]
ep = 0
s, e = (0 if ep == 0 else int(ends[ep - 1])), int(ends[ep])
wrist_rgb = root["data/camera0_rgb"][s:e]             # (T, 224, 224, 3) uint8
left = root["data/mic_0"][s:e].reshape(-1)             # (T * 2400,) at 48 kHz
right = root["data/mic_1"][s:e].reshape(-1)
eef_pos = root["data/robot0_eef_pos"][s:e]             # (T, 3)
gripper = root["data/robot0_gripper_width"][s:e]       # (T, 1)
store.close()
```

## Previews

One training episode per task as a short video with its stereo sound (headphones recommended).

| Path in this repository | Task | Episode | Target side | Length | Size |
|---|---|---|---|---|---|
| `previews/c4_T1_ep41_preview.mp4` | T1 Ring-off | 41 | left | 14.5 s | 0.40 MB |
| `previews/c4_T2_ep50_preview.mp4` | T2 Covered phone | 50 | left | 22.9 s | 0.77 MB |
| `previews/c4_T3_ep24_preview.mp4` | T3 Call bell | 24 | right | 26.3 s | 0.83 MB |
| `previews/c4_T4_ep21_preview.mp4` | T4 Animal feeding | 21 | left (dog) | 26.0 s | 0.82 MB |

Each video plays at 1x speed: wrist camera (left half) and external camera (right half) at 10 fps over the L/R 1–12 kHz level of the episode, H.264 448×288, with stereo AAC audio (48 kHz, 128 kbps); the left channel is `mic_0` and the right is `mic_1`, the 50 ms per-step windows joined in order (they do not overlap) and scaled by one gain for both channels, so the level difference between the channels is kept and nothing clips.

## Distractor sounds

In our experiments, distractor sounds from ESC-50 are mixed into the audio as augmentation during training. ESC-50 is not included in this repository. Get it from the original source: https://github.com/karolpiczak/ESC-50

## License

This dataset is released under the [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license. ESC-50, used only for distractor augmentation, is not included and keeps its own license.

	---
	license: cc-by-4.0
	pretty_name: Egocentric stereo-audio robot manipulation demonstrations (T1-T4)
	task_categories:
	- robotics
	tags:
	- robotics
	- audio-visual
	- stereo-audio
	- sound-source-localization
	- imitation-learning
	- manipulation
	- real-robot
	- zarr
	---

	# Egocentric stereo-audio robot manipulation demonstrations (T1–T4)

	Real-robot teleoperated demonstrations in which a robot arm uses a stereo microphone pair and an RGB camera on its wrist to find a sound source that starts out of the camera's view and then manipulates it.

	- 4 tasks, 311 episodes, 136,954 steps (about 114 minutes) at 20 Hz.
	- Per step: wrist RGB (policy input), wrist depth, external RGB and depth, 50 ms of stereo audio at 48 kHz, end-effector pose and gripper width.
	- Robot: Franka Research 3, teleoperated with a gamepad. Cameras: Intel RealSense RGB-D.

	## Files

	\| Path in this repository \| Task \| Original dataset name \| Size \|
	\|---\|---\|---\|---\|
	\| `T1_ring_off/replay_buffer.zarr.zip` \| T1 Ring-off \| `ral_v3_timer_ring_off_out_sight` \| 1.50 GB \|
	\| `T2_covered_phone/replay_buffer.zarr.zip` \| T2 Covered phone \| `rss_coverd_iphone` \| 2.74 GB \|
	\| `T3_call_bell/replay_buffer.zarr.zip` \| T3 Call bell \| `rss_short_time_bell` \| 2.44 GB \|
	\| `T4_animal_feeding/replay_buffer.zarr.zip` \| T4 Animal feeding \| `rss_dog_cat_duck` \| 3.02 GB \|

	Sizes are in decimal GB (10^9 bytes).

	## Tasks

	- T1 Ring-off. A kitchen timer rings somewhere on a semicircular workspace around the robot, out of the wrist camera's view at the start. The robot turns toward the sound and presses the timer to stop it.
	- T2 Covered phone. A phone rings under one of several identical towels. The robot finds the towel the sound comes from and pulls it off to uncover the phone.
	- T3 Call bell. A call bell rings briefly. The robot goes to the bell, picks up the fork lying next to it and puts the fork into the cutlery tray.
	- T4 Animal feeding. A speaker next to a bowl plays a dog, cat or duck call. The robot picks food from three trays at fixed positions (the animal call decides which tray) and drops it into the bowl (the direction of the call decides where).

	\| Task \| Original name \| Episodes \| Steps \| Episode length, mean (min–max) \| Left / Front / Right \| Held-out episodes (split seed) \| Recorded \|
	\|---\|---\|---\|---\|---\|---\|---\|---\|
	\| T1 \| `ral_v3_timer_ring_off_out_sight` \| 86 \| 24,332 \| 14.15 s (10.25–17.30 s) \| 43 / 3 / 40 \| 6, 7, 17, 35, 52, 59, 61, 71, 81 (seed 42) \| 2026-09-09, 22:59–23:48 KST \|
	\| T2 \| `rss_coverd_iphone` \| 80 \| 36,431 \| 22.77 s (19.40–27.60 s) \| 39 / 3 / 38 \| 3, 16, 18, 26, 51, 52, 62, 67 (seed 15) \| 2026-10-06, 20:37–22:14 KST \|
	\| T3 \| `rss_short_time_bell` \| 70 \| 36,597 \| 26.14 s (21.20–34.75 s) \| 32 / 4 / 34 \| 5, 6, 29, 43, 50, 59, 67 (seed 42) \| on or before 2026-10-06 (archive written 17:44 KST) \|
	\| T4 \| `rss_dog_cat_duck` \| 75 \| 39,594 \| 26.40 s (20.75–35.45 s) \| 36 / 3 / 36 \| 6, 8, 18, 29, 45, 47, 57, 72 (seed 44) \| 2026-10-07, 15:42–17:56 KST \|

	T4 class blocks (episodes are stored in recording order, 25 per class):

	\| T4 class \| Episodes \| Left / Front / Right \|
	\|---\|---\|---\|
	\| dog \| 0–24 \| 12 / 1 / 12 \|
	\| cat \| 25–49 \| 12 / 1 / 12 \|
	\| duck \| 50–74 \| 12 / 1 / 12 \|

	Notes on the table:

	- Left / Front / Right is the azimuth of the target point seen from the end-effector position at the first step, in the robot base frame (+x forward, +y left). \|azimuth\| ≤ 15° counts as front. The target point is the final end-effector position (T1, press), the first gripper close (T2 towel, T3 fork) or the first gripper re-open (T4, drop into the bowl).
	- Held-out episodes are the episodes we held out for offline evaluation. They come from the training code's split function with `val_ratio = 0.1`:
	```python
	import numpy as np
	n_val = min(max(1, round(n_episodes * 0.1)), n_episodes - 1)
	held_out = np.random.default_rng(seed=seed).choice(n_episodes, size=n_val, replace=False)
	```
	- Recorded comes from the per-step wall-clock timestamps of the original recordings. These timestamps are not stored in the T3 archive, so only the date the archive was written is known for T3.

	## Format

	Each file is a zipped zarr (v2) group in the [Diffusion Policy](https://github.com/real-stanford/diffusion_policy) / [UMI](https://github.com/real-stanford/universal_manipulation_interface) `ReplayBuffer` layout. All arrays under `data/` share the first axis: the steps of all episodes, concatenated in order at 20 Hz. `meta/episode_ends[i]` is the exclusive end index of episode `i`.

	\| Key \| Shape \| Dtype \| Content \|
	\|---\|---\|---\|---\|
	\| `data/camera0_rgb` \| (N, 224, 224, 3) \| uint8 \| Wrist RGB camera. This is the policy's image input. \|
	\| `data/camera0_depth` \| (N, 224, 224, 3) \| uint8 \| Wrist depth as a colorized image (see below). \|
	\| `data/camera1_rgb` \| (N, 224, 224, 3) \| uint8 \| External (third-person) RGB camera. Not a policy input. \|
	\| `data/camera1_depth` \| (N, 224, 224, 3) \| uint8 \| External depth as a colorized image. \|
	\| `data/mic_0` \| (N, 2400) \| float64 \| Left microphone waveform, 48 kHz, 2400 samples (50 ms) per step. \|
	\| `data/mic_1` \| (N, 2400) \| float64 \| Right microphone waveform, same layout. \|
	\| `data/robot0_eef_pos` \| (N, 3) \| float32 \| End-effector position (m) in the robot base frame. \|
	\| `data/robot0_eef_rot_axis_angle` \| (N, 3) \| float32 \| End-effector orientation as an axis-angle vector (rad). \|
	\| `data/robot0_gripper_width` \| (N, 1) \| float32 \| Gripper opening width (m). \|
	\| `meta/episode_ends` \| (E,) \| int64 \| Exclusive end index of each episode. \|

	- Images are stored one frame per chunk with JPEG-XL compression (`imagecodecs_jpegxl`), so reading them needs `imagecodecs` (see below). Other arrays use Blosc/LZ4.
	- Original streams were RGB and depth at 640×480, 30 fps per camera, stereo audio at 48 kHz, end-effector pose at 100 Hz and gripper width at 30 Hz. They were synchronized to 20 Hz, and the images were resized to 224×224.
	- The depth arrays are not metric depth. The recording code clips depth to 0.01–2.0 m, scales it linearly to 0–255 and applies the OpenCV VIRIDIS colormap. Pixels with no depth reading map to the colormap value of 0.
	- Audio: to get a continuous waveform, flatten the rows of an episode (`mic_0[s:e].reshape(-1)`). The microphone spacing used in our training configuration is 0.12 m (a configuration value, not a measurement).
	- There is no separate `action` array or timestamp array. Our training code builds the action from the proprioceptive arrays as `[eef_pos, eef_rot_axis_angle, gripper_width]` at the following steps, as in UMI.

	## Loading

	```python
	# pip install "zarr<3" imagecodecs huggingface_hub
	import zarr
	from huggingface_hub import hf_hub_download
	from imagecodecs.numcodecs import register_codecs

	register_codecs() # registers the JPEG-XL codec used by the camera arrays

	path = hf_hub_download(
	repo_id="joonsu0109/soundact_data",
	repo_type="dataset",
	filename="T1_ring_off/replay_buffer.zarr.zip",
	)
	store = zarr.ZipStore(path, mode="r")
	root = zarr.open(store, mode="r")

	ends = root["meta/episode_ends"][:]
	ep = 0
	s, e = (0 if ep == 0 else int(ends[ep - 1])), int(ends[ep])

	wrist_rgb = root["data/camera0_rgb"][s:e] # (T, 224, 224, 3) uint8
	left = root["data/mic_0"][s:e].reshape(-1) # (T * 2400,) at 48 kHz
	right = root["data/mic_1"][s:e].reshape(-1)
	eef_pos = root["data/robot0_eef_pos"][s:e] # (T, 3)
	gripper = root["data/robot0_gripper_width"][s:e] # (T, 1)

	store.close()
	```

	## Previews

	One training episode per task as a short video with its stereo sound (headphones recommended).

	\| Path in this repository \| Task \| Episode \| Target side \| Length \| Size \|
	\|---\|---\|---\|---\|---\|---\|
	\| `previews/c4_T1_ep41_preview.mp4` \| T1 Ring-off \| 41 \| left \| 14.5 s \| 0.40 MB \|
	\| `previews/c4_T2_ep50_preview.mp4` \| T2 Covered phone \| 50 \| left \| 22.9 s \| 0.77 MB \|
	\| `previews/c4_T3_ep24_preview.mp4` \| T3 Call bell \| 24 \| right \| 26.3 s \| 0.83 MB \|
	\| `previews/c4_T4_ep21_preview.mp4` \| T4 Animal feeding \| 21 \| left (dog) \| 26.0 s \| 0.82 MB \|

	Each video plays at 1x speed: wrist camera (left half) and external camera (right half) at 10 fps over the L/R 1–12 kHz level of the episode, H.264 448×288, with stereo AAC audio (48 kHz, 128 kbps); the left channel is `mic_0` and the right is `mic_1`, the 50 ms per-step windows joined in order (they do not overlap) and scaled by one gain for both channels, so the level difference between the channels is kept and nothing clips.

	## Distractor sounds

	In our experiments, distractor sounds from ESC-50 are mixed into the audio as augmentation during training. ESC-50 is not included in this repository. Get it from the original source: https://github.com/karolpiczak/ESC-50

	## License

	This dataset is released under the [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/) license. ESC-50, used only for distractor augmentation, is not included and keeps its own license.