datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
EgoLife_IMUimage-as-an-imu-finetuning
Image as an IMU: Real-world Finetuning Dataset
Official real-world finetuning dataset from Image as an IMU: Estimating Camera Motion from a Single Motion-Blurred Image (ICCV 2025 Oral).
[arXiv] [Webpage] [GitHub]
PIXL, University of Oxford
Jerred Chen, Ronald Clark
Dataset Details
This dataset consists of 32 sequences of real-world motion-blurred videos in various indoor scenes, captured using the iPhone 13 camera.
dataset_train_real-world.csv and… See the full description on the dataset page: https://huggingface.co/datasets/jerredchen00/image-as-an-imu-finetuning.1226_imu1_base_decay_corpus
IMU-1 Stage 2 Training Corpus (Decay Phase)
Pre-tokenized training data for Stage 2 (decay phase) of IMU-1, a sample-efficient 430M parameter language model.
Dataset Details
Property
Value
Tokens
~28B
Format
Memory-mapped NumPy (.npy)
Tokenizer
SmolLM2-360M
Vocab size
49,152
Data Sources
Stage 2 uses tighter quality filters compared to Stage 1:
DCLM-edu (higher threshold filtering)
FineWeb-edu
FineMath
Curated high-quality sources… See the full description on the dataset page: https://huggingface.co/datasets/thepowerfuldeez/1226_imu1_base_decay_corpus.1218_imu1_base_stable_corpus
IMU-1 Stage 1 Training Corpus (Stable Phase)
Pre-tokenized training data for Stage 1 (stable phase) of IMU-1, a sample-efficient 430M parameter language model.
Dataset Details
Property
Value
Tokens
~29B
Format
Memory-mapped NumPy (.npy)
Tokenizer
SmolLM2-360M
Vocab size
49,152
Data Sources
High-quality filtered web data including:
DCLM-edu (educational content filtered from DCLM)
FineWeb-edu
Curated web sources
Download… See the full description on the dataset page: https://huggingface.co/datasets/thepowerfuldeez/1218_imu1_base_stable_corpus.IMUG-Bench
IMUG-Bench: Benchmarking Unified Multimodal Models on Interleaved Understanding and Generation
**Lingyi Meng*1, Zecong Tang*†1, Haoran Li1, Tengju Ru1, Zhejun Cui1, Weitong Lian1, Qi Kang1, Hangshuo Cao1, Yichen Zhu1, Yechi Liu3, Kaixuan Wang2, Yu-Jie Yuan4, Chunwei Wang4, Yu Zhang‡1, Bo Dai‡2
1Zhejiang University 2The University of Hong Kong 3Institute of Automation, Chinese Academy of Sciences 4Huawei
*Equal contribution †Project leader … See the full description on the dataset page: https://huggingface.co/datasets/ccccEsion/IMUG-Bench.IMUWiFine
IMUWiFine: End-to-End Sequential Indoor Localization
Paper: End-to-End Sequential Indoor Localization Using Smartphone Inertial Sensors and WiFi
GitHub: https://github.com/IS2AI/IMUWiFine
Description: The IMUWiFine dataset comprises IMU and WiFi RSSI data readings recorded in sequential order with a fine spatiotemporal resolution. The
dataset was collected on the fourth, fifth, and sixth floors of the C4 building at the Nazarbayev University campus. The total covered area is over 9… See the full description on the dataset page: https://huggingface.co/datasets/issai/IMUWiFine.egocentric-gopro-rgb-imu
Hub Egocentric: GoPro RGB+IMU
19 egocentric human-manipulation clips captured on GoPro HERO13, with high-rate IMU (~200 Hz GPMF) delivered as CSV/Parquet/JSON sidecars plus a Foxglove MCAP recording.
Part of the Hub Egocentric Human Demonstrations Sample Set collection. Captured on GoPro HERO13. Egocentric, human-demonstration data (passive; no robot action stream). July 2026.
Dataset structure
Each clip is a top-level folder named #NN_... holding its media, a… See the full description on the dataset page: https://huggingface.co/datasets/Hubdata/egocentric-gopro-rgb-imu.IMU4DData
IMU4D Data
Processed motion / IMU training and evaluation data for IMU4D. The repo
mirrors the data/processed/ tree of the IMU4D_dev code base, so downloading
it into $IMU4D_DATA_ROOT/processed reproduces every default path used by the
training configs.
Dataset
Path
train / val / test
Size
MotionMillion + LINGO
motionmillion/v1/wds/
913,308 / 58,444 / 170,757
~111 GB
HiPHI
hiphi/v1/{wds,splits}/
13,037 / 500 / 500
~13 GB
OMOMO
omomo/v1/{wds,splits,samples}/
5,279… See the full description on the dataset page: https://huggingface.co/datasets/TianhangCheng7/IMU4DData.egocentric-iphone-rgb-imu
Hub Egocentric: iPhone RGB+IMU
12 egocentric human-manipulation clips captured on iPhone, with nominal 30 Hz CoreMotion + ARKit IMU/attitude on the shared media timeline, delivered as CSV/Parquet/JSON sidecars plus a Foxglove MCAP recording and per-clip camera intrinsics.
Across this 12-clip sample, IMU row count is 0–4 boundary rows lower than decoded video frame count (≤0.035%).
Part of the Hub Egocentric Human Demonstrations Sample Set collection. Captured on iPhone 13… See the full description on the dataset page: https://huggingface.co/datasets/Hubdata/egocentric-iphone-rgb-imu.aloha_mobile_put_egg_close_boxThis dataset was created using LeRobot.
Dataset Structure
meta/info.json:
{
"codebase_version": "v2.0",
"robot_type": "mobile_aloha",
"total_episodes": 202,
"total_frames": 148600,
"total_tasks": 1,
"total_videos": 0,
"total_chunks": 1,
"chunks_size": 1000,
"fps": 50,
"splits": {
"train": "0:202"
},
"data_path": "data/chunk-{episode_chunk:03d}/episode_{episode_index:06d}.parquet",
"video_path":… See the full description on the dataset page: https://huggingface.co/datasets/imumtozee/aloha_mobile_put_egg_close_box.runcam-feed-camera-egocentric-rgb-imu
RunCam Feed Camera - Egocentric RGB + IMU Sample Dataset
A small sample dataset captured with the RunCam Feed Camera for egocentric video and synchronized motion-sensor workflows.
Capture Device
Video: H.265 MP4, 1920x1080, 60 fps for V01-V06
Nominal video bitrate: 18 Mbps
Horizontal field of view: 126 degrees
Device weight: approximately 26 g
IMU: ICM-42607, 6-axis
IMU sampling rate: 800 Hz for the recordings in this sample
Firmware reported in the GCSV files:… See the full description on the dataset page: https://huggingface.co/datasets/RunCam/runcam-feed-camera-egocentric-rgb-imu.NFI_FARED_IMUThis is the README file for the dataset Netherlands Forensic Institute: Forensic Activity Recognition Dataset (NFI_FARED), published as a part of the paper "Hi-OSCAR: Hierarchical Open-set Classifier for Human Activity Recognition.". Two forms of data were collected: Digital Traces from iPhones worn on the subjects' bodies, and raw sensor signals from body-worn Inertial Measurement Units (IMUs). This dataset and README refers to the IMU data. The Digital Trace data is available here.
NFI_FARED… See the full description on the dataset page: https://huggingface.co/datasets/NetherlandsForensicInstitute/NFI_FARED_IMU.egocentric-video-imu-multimodal-sample-v1
Origin Data Lab — Egocentric Video + 6-DoF IMU Sample
This public technical sample demonstrates Origin Data Lab's capability to collect, structure, quality-check, and package real-world egocentric multimodal data for AI, robotics, and embodied-AI applications.
The sample contains egocentric video, task-level 6-DoF IMU data, sanitized metadata, and machine-measured sensor QC.
This repository is a limited public capability sample, not a complete production dataset.… See the full description on the dataset page: https://huggingface.co/datasets/origindatalab/egocentric-video-imu-multimodal-sample-v1.underwater-img-imu-sonar-datasetsLATAM-Egocentric-Residential-IMU
LATAM Egocentric Residential (with IMU)
Head-mounted, first-person video of everyday household chores recorded across Brazil, Argentina, Venezuela and Peru, each paired with a ~100 Hz accelerometer + gyroscope IMU stream.
The dataset targets embodied-AI and robotics research that needs real, unscripted human manipulation in cluttered domestic environments — not lab-staged demonstrations.
Preview: LATAM_OD_D_16 — gardening, outdoor, daytime, Argentina (30 s excerpt, downscaled… See the full description on the dataset page: https://huggingface.co/datasets/humyn-labs/LATAM-Egocentric-Residential-IMU.Human-Consistency
Human-Consistency
A deterministic 300-output sample from the successful Gemini-3-pro + TTS MMVC outputs. Each record includes the generated response audio, the exact rendered audio-judge prompt and criteria, and the judge's raw and parsed output.
Sampling
Seed: 20260828.
Population: 3,005 successful Gemini-3-pro + TTS outputs.
Allocation: 55 Emotional-Interaction, 234 Proactive-Care, 11 Safe-Companion-Behavior.
Proactive-Care: 214 paralinguistic, 10 semantic, 10… See the full description on the dataset page: https://huggingface.co/datasets/imuxsh/Human-Consistency.tracker-pov-imu
Eidon Tracker POV: IMU
24 Hz orientation and motion data from a seven-point IMU harness, paired with the egocentric
video in eidon-ai/tracker-pov.
One row per (recording, timestamp, body slot), roughly 780 million rows. Join to the video
metadata on recording_id.
This repo holds the sensor data only. There is no video here.
The release sits in three places:
Contents
Size
tracker-pov
the 13,451 MP4s and metadata.parquet
9.05 TB
this repo… See the full description on the dataset page: https://huggingface.co/datasets/eidon-ai/tracker-pov-imu.egocentric-wrist-view-camera-IMU-data
Egocentric Wrist-View Camera + IMU Data for Physical AI
Wrist-mounted camera + synchronized 6-axis IMU. The view a robot's wrist camera
actually has — close to the hand, moving with it, object filling the frame.
Free 30-episode sample. Task: folding and organizing cloth. LeRobot v3.0, loads in one line.
Does it help a policy? Measured.
We ran a controlled test on a public benchmark — DexGarmentLab
(NeurIPS 2025), cloth folding in Isaac Sim. A Diffusion Policy was… See the full description on the dataset page: https://huggingface.co/datasets/inhandplus/egocentric-wrist-view-camera-IMU-data.supermemory-vqa-imu-benchmark
SuperMemory-VQA
Inference-friendly v2 manifest for kfkas/supermemory-vqa-imu-benchmark.
This repository contains metadata and logical local asset references only. Full
source MP4, VRS, CSV, NPZ, or ZIP files are not redistributed here.
Tables
table
rows
purpose
benchmark
4,853
one row per evaluation question
assets
83
one row per synchronized source asset
evidence
7,014
one row per question-to-asset time relation
The Viewer uses the benchmark… See the full description on the dataset page: https://huggingface.co/datasets/kfkas/supermemory-vqa-imu-benchmark.lidar_imu_odometryIMUT-MCegolifeqa-imu-benchmark
EgoLifeQA
Inference-friendly v2 manifest for kfkas/egolifeqa-imu-benchmark.
This repository contains metadata and logical local asset references only. Full
source MP4, VRS, CSV, NPZ, or ZIP files are not redistributed here.
Tables
table
rows
purpose
benchmark
500
one row per evaluation question
assets
774
one row per synchronized source asset
evidence
1,048
one row per question-to-asset time relation
The Viewer uses the benchmark split. The… See the full description on the dataset page: https://huggingface.co/datasets/kfkas/egolifeqa-imu-benchmark.Voca-Human-Testing
Voca-Human-Testing
Voca-Human-Testing is a 252-record, audio-backed evaluation subset covering four
top-level companion capabilities and 14 second-level capabilities.
The subset was sampled from the four MMVC task files at approximately 10% per
second-level capability. Sampling jointly balanced available categorical metadata,
including source-dataset, subtype, label, style, tts_style, and voice.
Contents
Voca-Human-Testing.json: 252 complete benchmark records.… See the full description on the dataset page: https://huggingface.co/datasets/imuxsh/Voca-Human-Testing.UWB_IMU_GT_QDrone_Benchmark_DatasetFor additional details, please visit our website: https://benchmark.qdrone.ausmlab.com.
Q-Drone UWB Benchmark Dataset
Overview
We present the Q-Drone UWB Benchmark, a unique dataset derived from experiments conducted using the Q-Drone system—a UAV equipped with a UWB network at York University. This dataset encompasses data from five different sites, including an indoor environment, an open sports field, an area near a glass building, a semi-open tunnel, and beneath a… See the full description on the dataset page: https://huggingface.co/datasets/QDrone/UWB_IMU_GT_QDrone_Benchmark_Dataset.egor1-egomm-imu-benchmark
Ego-R1 Bench / EgoMM-EgoLife
Inference-friendly v2 manifest for kfkas/egor1-egomm-imu-benchmark.
This repository contains metadata and logical local asset references only. Full
source MP4, VRS, CSV, NPZ, or ZIP files are not redistributed here.
Tables
table
rows
purpose
benchmark
295
one row per evaluation question
assets
318
one row per synchronized source asset
evidence
333
one row per question-to-asset time relation
The Viewer uses the… See the full description on the dataset page: https://huggingface.co/datasets/kfkas/egor1-egomm-imu-benchmark.egomemreason-imu-benchmark
EgoMemReason
Inference-friendly v2 manifest for kfkas/egomemreason-imu-benchmark.
This repository contains metadata and logical local asset references only. Full
source MP4, VRS, CSV, NPZ, or ZIP files are not redistributed here.
Tables
table
rows
purpose
benchmark
500
one row per evaluation question
assets
306
one row per synchronized source asset
evidence
500
one row per question-to-asset time relation
The Viewer uses the benchmark split. The… See the full description on the dataset page: https://huggingface.co/datasets/kfkas/egomemreason-imu-benchmark.egogaze-vqa-egoexo4d-imu
EgoGazeVQA Ego-Exo4D subset
Inference-friendly v2 manifest for kfkas/egogaze-vqa-egoexo4d-imu.
This repository contains metadata and logical local asset references only. Full
source MP4, VRS, CSV, NPZ, or ZIP files are not redistributed here.
Tables
table
rows
purpose
benchmark
688
one row per evaluation question
assets
229
one row per synchronized source asset
evidence
688
one row per question-to-asset time relation
The Viewer uses the… See the full description on the dataset page: https://huggingface.co/datasets/kfkas/egogaze-vqa-egoexo4d-imu.multicam-dynamic-rig-imu
MultiCam · 动态 rig 立体 + IMU 数据集
8 视角双目 + 8 路 IMU,相机阵列本身在运动的一段同步采集。
为什么单独成库
同批 111 个采集会话里,用陀螺仪逐一筛查后,只有这一段的相机架真的动过:
陀螺峰值
本段 (ui_1785563256)
42.9 dps
第二名
1.40 dps
其余 63 段
0.3–0.9 dps(传感器噪声底)
差 30 倍,分离度极干净。其余会话都是"人动、架子静止",放在
DeDeHCL/multicam-4d-stereo。
规格
项
值
视角
8 台双目模组(DECXIN / Nori-3D,AR0234)
图像
每台 left+right,1920×1200,579 帧
帧率
60 Hz(ESP32 硬件触发,单源并联)
跨相机同步
0.525 µs
IMU
每台 1 个 ICM-42688P,601 Hz,6 轴
IMU 样本
每帧 11… See the full description on the dataset page: https://huggingface.co/datasets/HelloDeDe/multicam-dynamic-rig-imu.Imurka-datadolphinDolphin 🐬
https://erichartford.com/dolphin
Dataset details
This dataset is an attempt to replicate the results of Microsoft's Orca
Our dataset consists of:
~1 million of FLANv2 augmented with GPT-4 completions (flan1m-alpaca-uncensored.jsonl)
~3.5 million of FLANv2 augmented with GPT-3.5 completions (flan5m-alpaca-uncensored.jsonl)
We followed the submix and system prompt distribution outlined in the Orca paper. With a few exceptions. We included all 75k of CoT in the FLAN-1m… See the full description on the dataset page: https://huggingface.co/datasets/Imunlucky/dolphin.