Lab / Product

The motion capture system built for physical AI.

Phi9 MoCap Rig v1 captures full-body human motion synchronized with egocentric video. It is built for humanoid robotics data that can leave the lab, preserve the body, and arrive ready for training.

100 Hz Motion stream
<3 ms Sync drift
5 min Field setup
LeRobot Native export

REAL-TIME SYNC MULTI-MODAL CAPTURE IMU EGO-CENTRIC VIDEO DOMAIN-DIVERSE DATA PRETRAINING DATA POLICY TRAINING DATA WORLD-MODEL DATA RETARGETING TASK LABELS EVALUATION TRACES REAL-TIME SYNC MULTI-MODAL CAPTURE IMU EGO-CENTRIC VIDEO DOMAIN-DIVERSE DATA PRETRAINING DATA POLICY TRAINING DATA WORLD-MODEL DATA RETARGETING TASK LABELS EVALUATION TRACES

Problem

Video alone loses the body.

Egocentric datasets show what the camera saw, but not how the full kinematic chain moved. Arms-only capture works for narrow pick-and-place tasks. Humanoids need walking, reaching, bending, and coordination in the same example.

Solution

Wear it, calibrate, perform the task.

The rig pairs full-body motion with first-person video in real environments. No camera volume, no studio reset, and no separate alignment pass before the data can be used.

Core specifications

Built as capture infrastructure, not a demo prop.

Capture mode	Full-body or upper-body
Frequency	100 Hz
Egocentric video	Synchronized, head-mounted
Sync drift	<3 ms mocap + video
Environment	Homes, factories, outdoors, labs
Setup time	<5 minutes
Calibration	30 seconds
Battery	3 hours continuous
Output formats	BVH, MVNX, CSV, HDF5, LeRobot v2.1

Data quality

The useful part is what survives export.

Alignment

Frame-level sync

The hand reaching in egocentric video and the corresponding joint trajectory stay tied to the same timestamp.

Context

Metadata that trains

Task labels, subtask boundaries, object annotations, and quality scores travel with each episode.

Exports

Format-native output

Episodes move into LeRobot, HDF5, MuJoCo, Isaac Sim, Zarr, and ROS2 without rebuilding the pipeline.

Comparison

Full-body coverage with field deployment.

System	Body coverage	Frequency	Environment	Egocentric sync
Phi9 MoCap v1	Full body / upper body	100 Hz	Anywhere	Yes (<3 ms)
Xsens MVN	Full body	60-240 Hz	Studio-limited	Separate system
Rokoko Smartsuit	Full body	100 Hz	Studio-limited	Separate system
ALOHA	Arms only	50 Hz	Lab only	No
UMI	End-effector	Variable	Field	Yes, EE only
Egocentric video only	None	30-60 Hz	Anywhere	N/A

Pipeline

You capture. We process. You train.

01
Wear

Full-body suit plus a head-mounted camera, prepared by one operator.
02
Calibrate

Thirty seconds to register the body before natural task performance begins.
03
Capture

Motion, first-person video, and task traces are recorded together in the field.
04
Process

Quality scoring, segmentation, conversion, retargeting, and annotation run downstream.
05
Train

Clean episodes export into pretraining, policy training, simulation, and evaluation loops.

Use cases

Where the rig plugs into physical AI work.

Humanoid robotics

Whole-body demonstrations for locomotion, manipulation, and coordination.

Imitation learning

Behavior cloning, inverse RL, and diffusion policies with synchronized visual observation.

Physical AI development

Task-specific corpora for pretraining, mid-training, fine-tuning, and evaluation.

MoCap capture

Building capture systems for physical AI?

If you are working on motion capture, multi-modal data collection, or training pipelines for physical AI, write to us.

Write to

research@phi9.space

See the data