MicroAGI00: MicroAGI Egocentric Dataset (2025)

License: MicroAGI00 Open Use, No-Resale v1.0 (see LICENSE). No resale: You may not sell or paywall this dataset or derivative data. Trained models/outputs may be released under any terms.

What this is

MicroAGI00 is a large-scale egocentric RGB+D dataset of human household manipulation, aligned with the task style of the Stanford BEHAVIOR benchmark: https://behavior.stanford.edu/challenge/index.html

It’s designed to be “robotics-ready” at the signal level: synchronized streams, clean packaging, strong QC, and consistent structure—so you can spend time modeling, not cleaning data.

Quick facts

• Modalities: synchronized RGB + 16-bit depth + IMU
• Resolution & rate (RGB): 1920×1080 @ 30 FPS (in MCAP)
• Depth: 16-bit, losslessly compressed inside MCAP
• Scale: ≈1,000,000 synchronized RGB frames and ≈1,000,000 depth frames (≈1M frame pairs)
• Container: .mcap (all signals)
• Previews: for a subset of sequences, .mp4 previews (annotated overlays / visualized depth for quick review)

Note: MP4 previews may be lower quality than MCAP due to compression and post-processing. Research use should read from MCAP.

What’s included per sequence

• One large MCAP file containing:

  • RGB frames (1080p/30 fps)
  • 16-bit depth stream (lossless compression)
  • IMU data (as available)

• MP4 preview videos (subset of sequences):

  • RGB preview (for quick visual QA)
  • Visualized depth preview (for quick visual QA)

Labels / annotations https://www.youtube.com/watch?v=4-RVKBj2bcw

The base MicroAGI00 release is primarily raw synchronized signals (RGB-D-IMU) and does not ship with full-coverage labels.

If you’ve seen demo videos with overlays: those demonstrate what MicroAGI can produce as an add-on (see below), not what is universally present in the base dump.

Data quality and QC philosophy

MicroAGI00 is built around trustworthy signal integrity:

• Tight RGB↔Depth synchronization checks

• Automated detection and scoring of:

  • frame drops / time discontinuities
  • motion blur / exposure failures
  • depth sanity (range/invalid ratios), compression integrity
  • IMU continuity where available

• Consistent trimming and packaging, with sequence-level quality ratings to support filtering (e.g., “clean only” training vs. “wild” robustness training)

Diversity and covariate-shift robustness

MicroAGI data is captured across Europe and Asia, intentionally spanning environments that create real-world distribution shift:

• different homes, layouts, lighting regimes, materials
• different hands/skins, tool choices, cultural cooking/object priors
• varied camera motions and operator styles

This is meant to be covariate-shift resilient data for models that need to generalize.

Optional derived signals (available on request)

If you want more than raw RGB-D-IMU, MicroAGI can deliver derived outputs on top of the same sequences (or on newly captured data), such as:

• Ego-motion / trajectories (VIO-style)
• SLAM reconstructions (maps, trajectories, keyframes)
• Accurate body pose estimation
• State-of-the-art 3D hand landmarks (true 3D, not just 2D reprojections)
• Additional QA layers and consistency checks tailored to your training setup

These are provided as a service deliverable (and can be scoped to subsets / key frames / full coverage), depending on your needs.

Data access and structure

• Each top-level sample folder typically contains:

  • an MCAP “raw dump” folder
  • an MCAP “processed/curated” folder (when applicable)
  • an mp4/ previews folder (when available)

All authoritative signals are inside the MCAP. Use MP4s for fast browsing only.

Getting started

• Inspect an MCAP: mcap info your_sequence.mcap
• Extract messages: mcap cat --topics <topic> your_sequence.mcap > out.bin
• Python readers: pip install mcap (see the MCAP Python docs) or any MCAP-compatible tooling.

Typical topics include RGB, depth, IMU, and any additional channels you may have requested.

Intended uses

• Policy and skill learning (robotics / VLA)
• Action detection and segmentation
• Hand/pose estimation and grasp analysis (raw or with add-ons)
• Depth-based reconstruction, SLAM, scene understanding
• World-model pre/post training
• Robustness testing under real distribution shift

Data rights, consent, and licensing options

All capture is legally consented, with data rights documentation attached. Depending on the engagement, rights can be structured as:

• non-exclusive usage rights (typical dataset access), or
• exclusive rights for specific task scopes / environments / cohorts (custom programs)

Services and custom data

MicroAGI provides on-demand:

• New data capture via our operator network (Europe + Asia)
• ML-enhanced derived signals (ego-motion, pose, hands, SLAM)
• Real-to-Sim pipelines and robotics-ready packaging
• Custom QC gates to match your training/eval stack

Typical lead times: under two weeks (up to four weeks for large jobs).

How to order more

Email [email protected] with:

• Task description
• Desired hours or frame counts
• Target environment constraints (if any)
• Rights preference (exclusive / non-exclusive)
• Proposed price

We reply within one business day with lead time and final pricing.

Questions: [email protected]

License

This dataset is released under the MicroAGI00 Open Use, No-Resale License v1.0 (custom). See LICENSE. Redistribution must be free-of-charge under the same license.

Required credit: "This work uses the MicroAGI00 dataset (MicroAGI, 2025)."

Attribution reminder

Public uses of the Dataset or Derivative Data must include the credit line above in a reasonable location for the medium (papers, repos, product docs, dataset pages, demo descriptions). Attribution is appreciated but not required for Trained Models or Outputs.