SABER-10K

A real-world egocentric dataset of retail manipulation skills collected by DreamVu, formatted for LeRobot (v2.0).

The dataset covers retail environment navigation and manipulation recorded from a first-person (ego-view) perspective across three data streams. All episodes have passed a human deface quality-control audit.

Dataset Summary

Stream	Embodiment	Episodes	Frames	Tasks	Action Dim	Video Resolution
SABER-stream1	Egocentric (latent)	5,000	1,438,845	4,260	4	640×480
SABER-stream2	Humanoid (72-DOF)	200	38,204	200	72	640×360
SABER-stream3	Dexterous hands (36-DOF)	4,800	1,377,496	4,800	36	640×360
Total		10,000	2,854,545

Streams

SABER-stream1

Egocentric retail navigation with actions encoded using LAPA (Latent Action Pretraining from Videos), a codebook-based action quantization model trained on large-scale egocentric video. Each frame has 4 discrete latent action codes (codebook size 8).

• FPS: ~30 (29.97)
• Video: observation.images.ego_view — 640×480, H.264
• Action: float32[4] — LAPA latent codes (latent_0 … latent_3)
• State: float32[1] — dummy placeholder
• Tasks: 4,260 unique natural-language task descriptions

SABER-stream2

Full-body humanoid retargeted data in retail environments. State and action are 72-DOF joint configurations covering the entire body (legs, waist, arms, wrists, fingers, root pose and EEF poses) of a unitree g1 robot.

• FPS: ~30 (29.97)
• Video: observation.images.ego_view — 640×360, MP4V
• Action: float32[72] — full-body joint targets (root pose + 65 joints)
• State: float32[72] — full-body joint positions
• Tasks: 200 unique natural-language task descriptions

SABER-stream3

Dexterous hand manipulation in retail environments. State and action cover both hand poses and finger joint angles (36-DOF) of an inspire 5 finger gripper.

• FPS: ~30 (29.97)
• Video: observation.images.ego_view — 640×360, MP4V
• Action: float32[36] — hand pose + finger joints (left + right)
• State: float32[36] — hand pose + finger joint positions
• Tasks: 4,800 unique natural-language task descriptions

Parquet Schema

SABER-stream1

Column	Type	Description
index	int64	Global frame index across all episodes
episode_index	int64	Episode index within the stream
timestamp	float32	Time in seconds from episode start
task_index	int64	Index into meta/tasks.jsonl
observation.state	float32[1]	Dummy state placeholder
action	float32[4]	LAPA latent action codes
next.reward	float32	Reward signal
next.done	bool	Episode termination flag
annotation.human.action.instruction	int64	Task index (mirrors task_index)
annotation.human.validity	int64	Human validity annotation

SABER-stream2 / SABER-stream3

Column	Type	Description
index	int64	Global frame index across all episodes
episode_index	int64	Episode index within the stream
frame_index	int64	Frame index within the episode
timestamp	float32	Time in seconds from episode start
task_index	int64	Index into meta/tasks.jsonl
observation.state	float32[72 or 36]	Joint state vector
action	float32[72 or 36]	Joint action vector

Dataset Structure

Each stream follows the LeRobot v2.0 layout:

SABER-stream{N}/
├── meta/
│   ├── info.json               # Dataset metadata (fps, features, totals)
│   ├── episodes.jsonl          # Per-episode metadata (length, task indices)
│   ├── tasks.jsonl             # Task index → natural-language description
│   ├── stats.json              # Dataset-wide feature statistics
│   ├── modality.json           # Modality configuration
│   └── episode_mapping.csv     # New episode index → source episode index
├── data/
│   └── chunk-{NNN}/
│       ├── episode_000000.parquet
│       └── ...
└── videos/
    └── chunk-{NNN}/
        └── observation.images.ego_view/
            ├── episode_000000.mp4
            └── ...

Usage

With LeRobot

pip install lerobot

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="DreamVu/SABER-10K",
    root="SABER-stream1",
)

print(f"Episodes: {dataset.num_episodes}")
print(f"Frames:   {len(dataset)}")

frame = dataset[0]
# dict_keys(['observation.images.ego_view', 'action', 'timestamp', ...])

With HuggingFace datasets

from datasets import load_dataset

ds = load_dataset("DreamVu/SABER-10K", name="SABER-stream1", split="train")
print(ds)

Download a single stream

huggingface-cli download DreamVu/SABER-10K \
    --repo-type dataset \
    --include "SABER-stream1/**" \
    --local-dir ./SABER-10K

License

This dataset is released under CC BY-NC 4.0. It is intended for non-commercial research use only.