Sensor · Egocentric data
Egocentric SLAM & Visual-Inertial Datasets
An egocentric SLAM dataset pairs head-mounted camera video with time-synchronized inertial (IMU) streams and ground-truth camera trajectories, so a model can learn head-mounted visual-inertial odometry, localization, and mapping. This is head-mounted, human-worn SLAM — not vehicle-, drone-, or ground-robot-mounted odometry. The trajectory-complete open corpora (Aria Everyday Activities, Aria Digital Twin, Ego-Exo4D, Nymeria) are non-commercial or signed-license, so calibrated, pose-annotated custom capture is the route to commercially-usable VIO training data.
Quick facts
- Resolution
- 1080p @ 30 fps baseline; global-shutter and stereo preferred for metric scale and fast motion
- Field of view
- ≥120° horizontal so features stay in view through head turns; stereo optional for metric depth
- Mount
- Head-mounted glasses/rig with factory- or field-validated camera+IMU extrinsics — never chest or handheld, which decouple the IMU from the camera's motion
- Sensors
- RGB (baseline), Calibrated head IMU (accelerometer + gyroscope), Stereo (optional, for metric scale), Ground-truth 6-DoF camera trajectory / pose
- Labels
- Per-frame 6-DoF camera trajectory (position + orientation); Loop-closure / revisit markers on the trajectory; Camera-IMU calibration parameters (intrinsics + extrinsics)
- Volume
- 40–160 accepted hours per localization/mapping program
Key papers
Hard citations for the claims above. Each entry pairs a specific number with the paper that reports it.
EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video
829 hours, 194 tasks. EgoDex pairs 829 hours of egocentric video across 194 tabletop tasks with 3D hand and finger tracking captured on Apple Vision Pro — the largest and most diverse dexterous human-manipulation dataset to date.
Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives
1,286 hours, 740 participants. Ego-Exo4D pairs simultaneously-captured egocentric and exocentric video of skilled activity from 740 participants across 13 cities — 1,286 hours with multichannel audio, eye gaze, 3D point clouds, camera poses, and IMU.
The Monado SLAM Dataset for Egocentric Visual-Inertial Tracking
CC BY 4.0 head-mounted VIO. The Monado SLAM Dataset is one of the few genuinely head-mounted visual-inertial (VIO/SLAM) benchmarks — real VR-headset sequences covering high-intensity motion, dynamic occlusion, long tracking sessions and low-texture areas that existing datasets cover poorly — released under a permissive CC BY 4.0 license.
What SLAM & visual-inertial data contains
An egocentric SLAM dataset pairs head-mounted camera video with time-synchronized inertial (IMU) streams and ground-truth camera trajectories, so a model can learn head-mounted visual-inertial odometry, localization, and mapping. This is head-mounted, human-worn SLAM — not vehicle-, drone-, or ground-robot-mounted odometry. The trajectory-complete open corpora (Aria Everyday Activities, Aria Digital Twin, Ego-Exo4D, Nymeria) are non-commercial or signed-license, so calibrated, pose-annotated custom capture is the route to commercially-usable VIO training data.
The capture settings this covers:
- Walking a full loop through a building and returning to the exact start point — the revisit a loop-closure model has to recognize and snap the map back onto.
- Multi-room traversal: kitchen to hallway to stairwell to bedroom, where scale, texture, and lighting change at every doorway and the tracker has to hold pose through each transition.
- Re-entering the same corridor from the opposite direction, so the map must match features seen from a reversed viewpoint the camera never had head-on.
- Fast head turns and whip pans that blur the RGB frames and force the IMU to carry the trajectory until visual tracking re-locks.
- Dynamic scenes — people and carts moving through the frame while the wearer localizes against the static structure, the case that breaks naive feature matching.
- Long-horizon continuous capture (10–20 minutes) where odometry drift accumulates and only loop closure or a ground-truth trajectory can bound the error.
Why robotics and AI labs need SLAM & visual-inertial data
NVIDIA's Cosmos world-and-action-model work shows frontier labs treating large-scale first-person video as core training fuel for spatial and physical reasoning — and a world model that predicts what the camera sees next needs the camera's own trajectory, not just its pixels, which is exactly what SLAM/VIO data supplies. [1]
Apple's EgoDex pretrains dexterous manipulation from large-scale egocentric human video, and the same head-mounted RGB-plus-IMU capture that feeds it is what a visual-inertial SLAM stack ingests for camera-pose ground truth. [2]
EgoScale shows dexterous-manipulation performance scaling with the volume and diversity of egocentric human data, the same scaling argument that makes a large trajectory-annotated SLAM corpus worth capturing. [3]
Capture and delivery spec
Every SLAM & visual-inertial capture program runs to an explicit spec so the footage is training-ready on delivery rather than after a re-shoot. The baseline below is tuned per program; sensors, labels, and volume scale with the buyer's model.
| Spec | Detail |
|---|---|
| Resolution | 1080p @ 30 fps baseline; global-shutter and stereo preferred for metric scale and fast motion |
| Frame rate | 30 fps video baseline; IMU at 800 Hz–1 kHz, hardware-timestamped to frames |
| Field of view | ≥120° horizontal so features stay in view through head turns; stereo optional for metric depth |
| Mount | Head-mounted glasses/rig with factory- or field-validated camera+IMU extrinsics — never chest or handheld, which decouple the IMU from the camera's motion |
| Sensors | RGB (baseline), Calibrated head IMU (accelerometer + gyroscope), Stereo (optional, for metric scale), Ground-truth 6-DoF camera trajectory / pose |
| Labels | Per-frame 6-DoF camera trajectory (position + orientation); Loop-closure / revisit markers on the trajectory; Camera-IMU calibration parameters (intrinsics + extrinsics); Semi-dense point cloud / sparse map where captured; Keyframe timestamps and IMU-frame sync offsets |
| QA gates | Camera-IMU calibration validated per device before a batch is accepted; Trajectory ground-truth coverage and drift bound checked per clip; Field-of-view and horizon stability within threshold; Per-clip consent artifact attached (whole rooms and bystanders in frame) |
| Delivery | H.265 video + time-synchronized IMU + per-clip trajectory JSON (poses, keyframes, calibration), Hugging Face-streamable, with a consent artifact on every clip |
| Volume | 40–160 accepted hours per localization/mapping program |
Open SLAM & visual-inertial datasets
The 5 open corpora most relevant to SLAM & visual-inertial are compared below on scale, sensors, license, commercial use, and the gap each leaves for a buyer. None of them are cleanly licensed for commercial model training — which is the whole reason custom capture exists.
| Dataset | Size / scale | Sensors | License | Commercial use | Gap |
|---|---|---|---|---|---|
| Aria Everyday Activities | ≈7.3 h everyday capture · closed-loop trajectories + semi-dense point clouds (MPS) | RGB + SLAM cameras + dual IMU + eye-tracking; MPS trajectories & point clouds | Aria research/dataset license (non-commercial) | No | The trajectory + point-cloud quality is a gold standard, but the non-commercial license bars any shipping product and the scenes are a handful of Aria-team apartments, not your environments. |
| Aria Digital Twin | ≈200 instrumented sequences · GT 6-DoF device + object poses, depth, segmentation | RGB + IMU + simulated-twin ground-truth poses/depth | Aria research/dataset license (non-commercial) | No | Perfect synthetic-twin ground truth, but scripted instrumented rooms and non-commercial terms — the poses describe a fixed lab scene, not your building's loop topology. |
| Ego-Exo4D | ≈1,286 h across 13 cities · Aria SLAM/pose (MPS) streams, ego + exo | RGB + IMU + gaze + SLAM/pose streams | Ego-Exo4D signed data-use agreement | Conditional | Huge and pose-rich, but the signed agreement restricts commercial training and you inherit its skilled-activity task mix rather than the revisit-dense loops a SLAM model needs. |
| Nymeria | ≈300 h daily activity · Aria trajectories + full-body mocap | RGB + IMU + full-body motion capture + device trajectories | Aria research/dataset license (non-commercial) | No | Pairs trajectory with body motion, but it is non-commercial and built for human-motion research, so the capture is not designed around loop closure or long-horizon drift. |
| Monado SLAM Dataset | Head-mounted visual-inertial sequences for XR SLAM (arXiv:2508.00088) | Head-mounted camera + IMU + ground-truth poses | Open research license (verify terms at source) | Conditional | A rare genuinely head-mounted VIO benchmark, but it is small and XR-headset-rigged — a benchmark to test against, not a training-scale corpus for your domain. |
Open datasets vs Truelabel custom capture
Every trajectory-complete egocentric corpus is off-limits commercially. Aria Everyday Activities, Aria Digital Twin, and Nymeria are all non-commercial, and Ego-Exo4D's Aria SLAM streams sit behind a signed data-use agreement. You can benchmark VIO on them; you cannot ship a localization model trained on them.
Open corpora hand you the trajectory Aria's machine-perception pipeline produced, on Aria's device, in Aria's scenes. Custom capture lets you specify the environments (your warehouse, your home layouts, your factory floor), the device class, and the loop density, then delivers per-frame 6-DoF trajectories plus camera-IMU calibration in your episode format.
Loop closure and long-horizon drift only get exercised if the capture deliberately revisits places. You cannot guarantee an open corpus contains the loop topology your SLAM model has to solve — custom capture scripts the loops, the reversed-direction revisits, and the multi-room traversals on purpose.
SLAM footage records whole rooms, so provenance is not optional. Custom capture ships a validated per-device camera-IMU calibration instead of a reverse-engineered factory profile, plus a per-clip consent chain for wearers and bystanders and the option of exclusivity, so the same trajectories are not also sold to the lab you are racing.
SLAM & visual-inertial: by the numbers
The figures below are specific to SLAM & visual-inertial egocentric data and anchor the comparisons above.
- Aria Everyday Activities ships closed-loop device trajectories plus semi-dense point clouds from the Aria machine-perception pipeline — the field's gold-standard egocentric SLAM ground truth, and non-commercial
- Ego-Exo4D provides Aria SLAM/pose streams across ≈1,286 hours and 13 cities, but only under a signed data-use agreement
- Monado SLAM Dataset (arXiv:2508.00088) is one of the few genuinely head-mounted visual-inertial benchmarks — small, XR-rigged, and built to test VIO rather than train it
- "SLAM dataset" runs ~30/mo at LOW competition with a 100% academic/GitHub SERP (DataForSEO + Serper, 2026-07-06) — no commercial page owns the head-mounted framing
How Truelabel captures SLAM & visual-inertial data
Truelabel runs SLAM & visual-inertial programs on a network of 20,000+ consented collectors across nine countries, capturing to your brief on a head-mounted rig. Every clip passes per-clip machine QA — head-mount stability, field of view, and hands-in-frame coverage — and ships with a signed wearer consent artifact and provenance manifest. A calibration pilot returns its first batch in days, then accepted batches scale to 40–160 accepted hours per localization/mapping program, delivered as H.265 video + time-synchronized IMU + per-clip trajectory JSON (poses, keyframes, calibration), Hugging Face-streamable, with a consent artifact on every clip. Go deeper via what egocentric data is, egocentric data licensing, semi-dense point clouds, world models, humanoid pretraining data, and physical AI data marketplace.
Related pages
Use these to move from category-level context into specific task, dataset, format, and comparison detail.
External references and source context
- Physical AI with World Foundation Models | NVIDIA Cosmos
NVIDIA Cosmos world/action-model work treats large-scale egocentric video as core training fuel for spatial and physical reasoning.
NVIDIA ↩ - EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video
EgoDex is a large-scale egocentric video corpus built to pretrain dexterous manipulation from first-person human video.
arXiv ↩ - EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data
EgoScale supports the claim that dexterous-manipulation performance scales with the volume and diversity of egocentric human data.
arXiv ↩ - Project Aria Hardware Specifications
Project Aria devices carry SLAM cameras and IMUs whose machine-perception pipeline produces closed-loop trajectories and semi-dense point clouds.
Meta / Project Aria - Aria Everyday Activities (AEA)
Aria Everyday Activities is a non-commercial research release with closed-loop trajectories and semi-dense point clouds.
Meta / Project Aria - Aria Digital Twin (ADT)
Aria Digital Twin provides ground-truth 6-DoF device and object poses, depth, and segmentation under a non-commercial research license.
Meta / Project Aria - Ego-Exo4D project site
Ego-Exo4D provides Aria SLAM/pose streams across roughly 1,286 hours and 13 cities under a signed data-use agreement.
ego-exo4d-data.org - Nymeria: egocentric full-body motion dataset
Nymeria pairs Aria device trajectories with full-body motion capture across roughly 300 hours under a non-commercial research license.
Meta / Project Aria - Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot Transfer
Figure's Project Go-Big centers egocentric human video for humanoid pretraining, which depends on consistent ego-pose for spatial grounding.
Figure - Humanoid data: 10 Things That Matter in AI Right Now | MIT Technology Review
MIT Technology Review documents the real-world-data bottleneck constraining humanoid robots, of which spatial/localization data is a part.
MIT Technology Review
FAQ
What is an egocentric SLAM dataset?
It is head-mounted, human-worn camera video paired with time-synchronized IMU streams and ground-truth camera trajectories, so a model can learn visual-inertial odometry, localization, and mapping from the viewpoint a person (or a humanoid) actually moves through the world. It is distinct from car, drone, or ground-robot SLAM: the motion is human gait and head turns, and the value is in the paired IMU and the 6-DoF trajectory, not the RGB alone.
Do clips ship with camera trajectories and poses, or just raw video?
With trajectories, to spec. The baseline delivers per-frame 6-DoF camera pose, keyframe timestamps, IMU-frame sync offsets, and the camera-IMU calibration used to produce them; where captured, we also ship a semi-dense point cloud or sparse map. Raw-video-only is the failure mode of scraped footage — a SLAM model needs the pose target, which is exactly what the Aria machine-perception pipeline provides for the open corpora.
Why can't I just use Aria's open SLAM data (AEA, ADT, Nymeria, Ego-Exo4D)?
Because the trajectory-complete Aria corpora are legally untouchable for commercial work. Aria Everyday Activities, Aria Digital Twin, and Nymeria are non-commercial research releases, and Ego-Exo4D's SLAM streams require a signed data-use agreement. They are excellent for prototyping and for benchmarking a VIO pipeline; a shipping localization model needs pose-annotated footage you actually hold the rights to.
Which devices provide calibrated IMU + camera rigs at collector scale?
Head-mounted rigs with validated camera-IMU extrinsics — glasses-class devices or GoPro-style head mounts with a per-device calibration pass. The point is that the IMU is rigidly coupled to the camera and the extrinsics are known and checked per unit; chest and handheld mounts break that coupling. Aria Gen 2's calibrated SLAM cameras and dual IMUs are why wearable VIO capture is now practical at scale.
Can you capture long-horizon loops and revisits for loop-closure training?
Yes — that is the point of scripting the capture. We define the routes so the wearer closes loops, re-enters corridors from the opposite direction, and traverses multiple rooms in one continuous take, then hold the take long enough that drift accumulates and loop closure has something to correct. Open corpora contain whatever loops their tasks happened to produce; a targeted capture guarantees the topology your model needs.
How is consent handled when SLAM capture records whole rooms and bystanders?
Every clip carries a consent artifact tied to the wearer, bystander handling is defined up front, and PII is treated per your retention rules. Whole-room mapping footage is higher-exposure than a hands-only manipulation clip, so the provenance chain — which the open corpora rarely document at the clip level — is part of delivery, not an afterthought.
Looking for SLAM dataset?
Specify modality, task, environment, rights, and delivery format. Truelabel matches you with vetted capture partners and helps scope consent artifacts and commercial licensing requirements before delivery.
Request egocentric SLAM data