Open Source RoboticsRobots, software & learning

EgoAllo

Software

Estimate human body and hand motion from Project Aria camera and SLAM recordings, with separate visualization, inference and training workflows. [1]

Installation & first result →Official websiteGetting started

For
Researchers preparing human motion from wearable-camera recordings [1]
Source rights
Some licenses are unresolved. Review licenses
Unknowns
operating system not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
License
MIT · license not fully assessed for model weights, datasets · Details by part

Requirements · Research details and dates

On this page

Overview

EgoAllo estimates human body pose, height and hand motion from wearable-camera images and SLAM poses. Its documented input workflow uses Project Aria recordings, making it relevant to researchers collecting human motion for later analysis or retargeting. Robot retargeting and control are separate work; this entry covers the human-motion estimation pipeline. [1]

Is this for you?

Known limitations
  • Viewer also requires CUDA and the original trajectory context.
  • External model and dataset rights are separate from repository MIT terms.
  • The README installs jaxls from an unpinned Git URL.

Editorial assessment from cited documentation; no runtime testing.

[1] [3] · Editorial assessment

What you need

Languages / runtimes
  • Python

[2]

Dependencies
  • Python >=3.12
  • torch 2.7.1
  • projectaria_tools 1.6.0
  • Separately obtained SMPL-H model
  • Inference: JAX and jaxls

[1] [2]

Compute for the workload
  • Both viewer and inference select a CUDA device; no workload-specific minimum GPU memory is established here.

Editorial assessment from cited documentation; no runtime testing.

[3] [4] · Editorial assessment

Installation & first result

Open the upstream getting-started guide [1]

  • Create a Python 3.12+ environment and install the pinned source.
  • Obtain the sample trajectories and extended SMPL-H model under their own terms.
  • Run python 4_visualize_outputs.py --search-root-dir ./egoallo_example_trajectories on a CUDA-capable environment.

Editorial assessment from cited documentation; no runtime testing.

[1] [3] · Editorial assessment

Installation

Install the pinned checkout with pip install -e .; acquire sample data and SMPL-H separately. Inference additionally requires JAX and jaxls.

[1]

Hardware, models & data

Inputs and outputs

Aria VRS recordings and MPS trajectories/points feed inference; motion parameters are saved as NPZ files.

[1] [4]

Project notes

Start by viewing the supplied outputs

Check out commit 0c20ae8c8d48be4ff682803ea523bcb00a6a708d. Use Python 3.12 or newer and install that checkout with pip install -e .. The package pins PyTorch 2.7.1 and Project Aria tools 1.6.0. Obtain the linked example trajectories and the extended SMPL-H model with 16 shape parameters; the default model path is ./data/smplh/neutral/model.npz. Then run python 4_visualize_outputs.py --search-root-dir ./egoallo_example_trajectories. [1] [2]

Viewing is not a lightweight CPU-only route: the script explicitly selects a CUDA device. It searches for egoallo_outputs/*.npz below the supplied root and also loads the trajectory's VRS recording, calibration and point cloud. Keep that directory context intact rather than copying an output file on its own. This source review does not establish a minimum GPU-memory requirement. [3]

Inference adds another dependency boundary

For inference, the README adds JAX with CUDA support, showing jax[cuda12]==0.6.1, and installs jaxls from its Git repository without a revision pin. Record the resolved jaxls commit when reproducing this setup; the current jaxls release state alone does not establish compatibility with EgoAllo. Download and unpack the linked checkpoint using the README script or manual link. The script expects ./egoallo_checkpoint_april13/checkpoints_3000000/ by default; use --checkpoint-dir for a different location. [4] The example command is python 3_aria_inference.py --traj-root ./egoallo_example_trajectories/coffeemachine. [1]

Own recordings need an Aria VRS file plus MPS closed_loop_trajectory.csv and semidense_points.csv.gz. HaMeR results and Aria wrist/palm tracking are optional additional inputs. The inference script also selects CUDA and saves estimated motion under egoallo_outputs. To generate HaMeR inputs, upstream points to the separate hamer_helper project. [1] [4]

Code access does not settle asset permissions

The repository code and documentation have MIT terms. The checkpoint and example trajectories are externally linked; their payloads and separate license scope were not verified here. SMPL-H and AMASS must be obtained separately under their own terms. Upstream explicitly says it cannot redistribute its preprocessed AMASS data because of licensing constraints, so retraining is not supplied by downloading the code alone. [1] [7]

This profile reviews a pinned commit snapshot, with no GitHub release listed. Source was preserved and inspected; installation, motion quality and downstream robot use have not been tested. [6]

What we checked

Source researched only; no installation, training, benchmark or physical hardware test performed. No affiliate links or reused media.

Not yet established

  • Requirements: operating system, host computer
  • Version-specific hardware compatibility (ecosystem tags alone do not establish it)

Updates

Upstream last updated
Sep 3, 2025 [6]
Latest release
Not established
Profile updated
Oct 17, 2026
Sources last checked
Oct 17, 2026
Next review
Nov 17, 2026
Reviewed revision
commit snapshot; no GitHub releases listed / 0c20ae8c8d48be4ff682803ea523bcb00a6a708d · software repository

Source files and licenses

Human motion estimation source is available; model and dataset access and rights remain separately scoped.

LayerAvailability / license
softwareavailable · MIT
Repository materials only; external components retain their own terms. [7]
model weightsavailable · License not fully assessed
Upstream supplies external asset links; only listing availability established, not payload acquisition or separate rights. SMPL-H and AMASS retain separate terms. [1]
datasetsavailable · License not fully assessed
Upstream supplies external asset links; only listing availability established, not payload acquisition or separate rights. SMPL-H and AMASS retain separate terms. [1]
documentationavailable · MIT
Repository materials only; external components retain their own terms. [7]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5

Change history