Stera-10M is an open egocentric multimodal dataset for embodied AI, robotics, world models, and spatial intelligence, captured end-to-end on commodity iPhone Pro hardware through the open Stera platform.

It contains 200 hours of synchronized first-person recordings across 500+ sessions from 20 contributors in 20+ unique environments, with 10 million RGB frames, LiDAR depth, ARKit 6-DoF camera pose, IMU, 21-joint MANO two-hand mocap, room mesh geometry, and hierarchical action language.

Stera-10M is, to our knowledge, the first open egocentric dataset combining hour-plus continuous capture with depth, 6-DoF pose, and dense hand annotations collected and distributed on commodity hardware, which a researcher already owns.

Stera-10M is built for training and evaluating models that learn from real human work as a continuous, geometrically grounded, multimodal signal, not as short, isolated clips.

It is designed to support research and product development in:

Vision-language-action (VLA) pretraining
Embodied AI and world modeling
Robot learning from human demonstration
Long-horizon manipulation and planning
Egocentric perception
Action understanding at multiple temporal scales
Hand-object interaction
3D / 4D scene and motion understanding
Real-to-sim and sim-to-real pipelines

Check out the Stera Capture iOS app to record your own sessions in the same format. Check out the Stera SDK for the open pipeline used to process every recording in this dataset.

Dataset Statistics

Statistic	Value
Total recording duration	200 h
Sessions	584
Mean session length	20.5 min
Longest continuous session	104 min
RGB frames	~10M
Atomic action spans	75,857
Episodes	9,922
Sub-goals	2,212
Total dataset size	1,600 GB

What makes Stera-10M different

Most existing egocentric datasets present a trade-off. Large datasets like Ego4D rely on heterogeneous capture hardware and lack consistent depth, 6-DoF pose, or hand annotations. High-fidelity datasets such as EgoExo4D and Aria Everyday Activities require gated, research-grade hardware (Project Aria) and involve short episode lengths. Production-quality manipulation datasets like HOI4D, HOT3D, and ARCTIC are limited to short clips in constrained settings.

Stera-10M is designed differently. It treats egocentric capture as a long-horizon, multimodal, fully open signal recorded on hardware anyone can buy. Each session includes:

what the wearer sees
how the camera moves through space (6-DoF pose)
how the hands move (21-joint MANO, anchored in a global frame)
what the depth geometry looks like (per-frame LiDAR depth and a session-level room mesh)
what the IMU measures
what task, sub-goal, episode, atomic action, and objects are involved (hierarchical instruction tree)

This makes Stera-10M especially useful for labs that need geometric grounding across hour-plus continuous activity, and for any researcher who wants to extend the corpus with their own captures using the same open stack.

Supported Tasks and Use Cases

Stera-10M can support a broad range of tasks, including but not limited to:

Vision-Language-Action model pretraining
Imitation learning and behavior cloning
Long-horizon manipulation policy learning
Egocentric action recognition
Temporal action localization and segmentation
Action captioning at multiple temporal scales
Task and sub-goal prediction
Hand-object interaction understanding
Human-to-robot motion retargeting
Hand pose estimation
Visual odometry and trajectory learning
SLAM and camera pose estimation
Visual-language pretraining
World model training
Real-to-sim environment reconstruction

Dataset Structure

Stera-10M is organized as a collection of sessions. Each session is a self-contained directory containing the egocentric RGB recording and a unified annotation.hdf5 file, the room mesh, calibrations, and rerun-sdk visualization recordings.

Session Layout

A typical session directory contains:

<session_id>/
├── rgb.mp4                   # 1280×720 @ 15 fps, H.264, PII encrypted
├── annotation.hdf5           # all per-frame and sparse annotations
├── mesh.ply                  # room mesh (~45k verts)
├── thumbnail.jpg
├── calibrations/
├── visualization.rrd         # full-session rerun recording
└── visualization_2min.rrd    # 2-min preview rerun recording

Modalities

Each session include the following modalities:

RGB video from a head-mounted iPhone Pro (1280×720 @ 15 fps)
LiDAR depth aligned to RGB (144×256, uint16 mm, per frame)
ARKit 6-DoF camera pose (rotation, translation, timestamp, per frame)
IMU (linear acceleration, angular velocity, orientation quaternion at ~100 Hz)
Two-hand mocap (21-joint MANO, anchored in the global frame, with MANO betas and finger rotations)
Room mesh (~45k vertices per session)
Session metadata (frame counts, duration, start/end times)
Hierarchical language captions at four temporal levels:
- session
- sub-goal
- episode
- atomic action

The Stera Platform

Stera-10M is the launch corpus for the Stera platform, an open, full-stack system for capturing, processing, evaluating, and exporting embodied AI data on commodity hardware. Every component used to produce this dataset is open and reproducible:

Stera Capture: an iOS app that turns an iPhone Pro into a synchronized RGB / depth / IMU / 6-DoF pose recorder, voice-controlled and writing to MCAP
Stera SDK: the processing pipeline used to generate every annotation in this dataset (trajectory, hand pose, hierarchical labels, upper-body tracking, mesh)
Stera Evaluate: the quality layer that scores tracking drift, hand-detection rate, label coverage, and instruction-tree integrity
Stera Export: converters to standard HDF5, PLY, RRDs, etc

Researchers who want to extend Stera-10M, capture in domains we have not covered, or build entirely new datasets in this format can do so using the same open stack.

Links:

Privacy, Ethics, and Consent

Because Stera-10M contains egocentric recordings of real-world human activity, privacy and consent are central considerations.

All contributors signed informed-consent agreements prior to recording. Recording was paused around non-consenting individuals. Incidentally captured faces are blurred via Stera SDK prior to release. Personally identifying or sensitive content is handled according to the dataset release policy.

Access

Stera-10M is released for research/commercial use and other non-commercial uses under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license with a carve out for WiLoR & MANO licences which are subject to separate licence terms.

Please make sure you understand the non-commercial & attribution carve outs as specified above by reading the complete licence. Licence here

Citation

@article{palanisamy2026mobileego,
  title   = {MobileEgo Anywhere: Open Infrastructure for long horizon egocentric data on commodity hardware},
  author  = {Palanisamy, Senthil and Anand, Abhishek and Rathore, Satpal Singh and Patnaik, Pratyush and Khatana, Shubhanshu},
  journal = {arXiv preprint arXiv:2605.05945},
  year    = {2026}
}

Please also credit WiLoR (Potamias et al., 2024), MANO (Romero et al., 2017), EgoBlur, and rerun-sdk.

Contact

FPV Labs - contact@fpvlabs.ai

Downloads last month: 15,147

Total file size:

1.64 TB

Paper for fpvlabs/stera-10m

MobileEgo Anywhere: Open Infrastructure for long horizon egocentric data on commodity hardware

Paper • 2605.05945 • Published 15 days ago • 9