The motion capture system built for physical AI.

The Problem with Current Robotics Data

Egocentric video doesn’t scale for physical AI.

You get what the camera sees. But you don’t get how the body moved. Joint angles. The full kinematic chain that makes manipulation possible.

Existing datasets capture 7-14 degrees of freedom — arms only. That works for pick-and-place. It doesn’t work for humanoids that need to walk, reach, bend, and coordinate their entire body.

High-fidelity dexterity data doesn’t exist at scale.

Motion capture systems require studios. Camera rigs. Hours of setup. They capture actors in controlled environments — not humans performing real tasks in real environments.

We built something different.

The Solution

Phi9 MoCap Rig v1 captures full-body human motion synchronized with egocentric video. No studio. No camera setup. Works in any environment.

Wear the suit. Attach the head-mounted camera. Perform tasks naturally. The system captures everything — what you saw, and how your body moved.

The result: Training data that scales. Full-body kinematics at 100 Hz, paired with first-person video, from real-world environments.

Core Specifications

Specification	Value
Capture Mode	Full-body or upper-body
Frequency	100 Hz
Egocentric Video	Synchronized, head-mounted
Sync Drift	<3ms (mocap + video)
Environment	Anywhere — homes, factories, outdoors
Setup Time	<5 minutes
Calibration	30 seconds
Battery	3 hours continuous
Output Formats	BVH, MVNX, CSV, HDF5, LeRobot v2.1

Why This Matters for Physical AI

Pre-Training

Foundation models need massive, diverse motion data. Our rigs deploy anywhere — capture thousands of hours of natural human behavior across environments. Not actors in studios. Real people doing real tasks.

Mid-Training

Domain adaptation requires task-specific demonstrations. Record manipulation sequences with full joint trajectories. Retarget to any humanoid platform.

Post-Training

Fine-tuning needs high-quality reference data. Per-episode quality metrics. Clean synchronization. Metadata that actually helps models learn.

Data Quality

Synchronized capture: Motion and video align at the frame level. When the hand reaches for an object in video, the corresponding joint trajectory is at the exact same timestamp. <3ms drift across 3-hour sessions.

Metadata-rich: Every episode ships with task labels, subtask boundaries, object annotations, and quality scores. The metadata that makes data usable for training.

Format-native: Export directly to LeRobot, HDF5, MuJoCo, Isaac Sim. Data formats that integrate with your existing pipeline — no conversion friction.

Comparison

System	Body Coverage	Frequency	Environment	Egocentric Sync
Phi9 MoCap v1	Full body / Upper body	100 Hz	Anywhere	Yes (<3ms)
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

The difference: We combine full-body motion capture with synchronized egocentric video, designed for field deployment, at a price point that scales.

Ease of Use

One operator. Five minutes. Any environment.

No camera calibration. No studio setup. No technical expertise required.

Put on the suit. Calibrate in 30 seconds. Start capturing.

This simplicity enables something important: anyone can contribute data. Not just trained operators in labs. Anyone wearing the suit, performing everyday tasks in their home, workplace, or any environment.

The result: Data diversity that reflects real-world complexity. Not sterile lab demonstrations.

Integration with Phi9 Data Pipeline

The rig isn’t standalone hardware. It’s the capture layer of the Phi9 data infrastructure.

Captured data flows directly into:

• Automated quality scoring
• Episode segmentation
• Format conversion (LeRobot, HDF5, Zarr, MuJoCo, Isaac Sim, ROS2)
• Retargeting to humanoid URDFs
• Language annotation pipelines

You capture. We process. You train.

Use Cases

Humanoid Robotics Training Full-body demonstrations for bipedal locomotion, manipulation, and whole-body coordination. Retarget to Figure, 1X, Unitree, Tesla, or any humanoid platform.

Imitation Learning Research Behavior cloning, inverse RL, diffusion policies — trained on complete joint trajectories with synchronized visual observation.

Physical AI Development Pre-training corpora, domain-specific mid-training, fine-tuning datasets. The data infrastructure for embodied intelligence.

From Capture to Training

1. Wear the rig — Full-body suit + head-mounted camera
2. Calibrate — 30 seconds
3. Perform tasks — Natural motion in real environments
4. Upload — Data flows into Phi9 pipeline
5. Export — Ready-to-train formats for your framework