Phi9 capture loop rendered as an archival robotics atlas with motion capture, sensor traces, robot hands, and evaluation fragments.

Phi9 — physical AI lab

Physical AI research, data systems, and deployable intelligence.

Phi9 turns real-world behavior into reusable training signal, evaluation systems, and deployable physical intelligence.

Explore the Lab See the Data
The loop

From capture to evaluation, the physical AI loop.

One loop. Four stages. The work is turning real-world behavior into signal, training against it, and measuring what survives contact with reality.

Archival atlas image showing motion capture, robotic hands, sensor traces, and synchronized physical AI capture artifacts.

First signal

Capture

The loop starts with wearable capture: full-body motion, egocentric video, IMU traces, task metadata, and live sync. The goal is not a studio recording; it is a reusable demonstration with timing, intent, and physical context still attached.

mocap suit -> ego video -> imu rail -> task trace

001 Capture
Archival atlas image showing motion capture, robotic hands, sensor traces, and synchronized physical AI capture artifacts.

First signal

Capture

The loop starts with wearable capture: full-body motion, egocentric video, IMU traces, task metadata, and live sync. The goal is not a studio recording; it is a reusable demonstration with timing, intent, and physical context still attached.

mocap suit -> ego video -> imu rail -> task trace

002 Multiply
Archival atlas image showing layered data catalogs, waveform strips, robot sequence panels, and simulation-ready data fragments.

Simulation expansion

Multiply

Each captured episode becomes a family of usable examples: segmented demonstrations, retargeted motion, synthetic variations, simulation rollouts, and data products that increase coverage without discarding the original task.

demo_01 -> segment -> retarget -> sim variants

003 Train
Archival atlas image showing robot manipulation sequences, policy traces, and training visualization panels.

Policy formation

Train

Training should look like aligned streams becoming action: motion state, first-person context, labels, reward curves, and policy rollouts compressed into one legible surface.

qpos + ego + imu + labels -> action chunks

004 Evaluate
Archival atlas image showing benchmark trajectories, evaluation curves, lunar terrain, and transfer proof panels.

Benchmarks and transfer

Evaluate

Evaluation is the benchmark layer: policy rollout, transfer test, failure trace, and deployment feedback compared in one place so the next capture and training pass know what to fix.

benchmark -> rollout -> failure trace -> next pass

  • 4

    Loop stages

  • 12+

    Modalities per capture

  • 25+

    Environment types

Featured system

The capture system behind the loop.

The Phi9 MoCap Rig records full-body motion and first-person video together, timed and labeled by task. The output is reusable training data instead of one-off recordings, so the same demonstration can move through training, fine-tuning, and deployment.

Open the MoCap Rig
Research questions

The bottlenecks we are actively working through.

These are not abstract themes. They are the constraints shaping the systems, experiments, and artifacts we are building now.

  1. 01

    Data that carries intent, not just observation.

    Most pipelines record visible motion but lose the task underneath it. We are working on capture that preserves action, context, and what the body was trying to achieve.

  2. 02

    Physical data is expensive.

    You cannot scrape physical behavior. Every demonstration needs a rig, a subject, a calibration, and a clean task. The work is making each capture travel further without losing signal.

  3. 03

    Benchmarks that predict real-world performance.

    A benchmark score means little if a policy falls apart on an unscripted task. We care about evaluation that predicts transfer, failure, and what survives outside the benchmark.

  4. 04

    One loop, not three stages.

    Capture, training, and evaluation still get treated as separate departments. We are trying to wire them into one visible loop so progress does not disappear between stages.

Research

Projects, notes, and technical progress from the lab.

Open work from the loop: systems, experiments, and technical questions published as they evolve.

Policy-based Deep RL: Lunar Landing
reinforcement-learning 4 min

Policy-based Deep RL: Lunar Landing

Pure Monte Carlo in an actor-critic setting — training a lunar lander with REINFORCE and analyzing the learned policy's decision-making.
The First Project: Hierarchical RL for Robotic Manipulation
reinforcement-learning

The First Project: Hierarchical RL for Robotic Manipulation

From sparse rewards to Gaussian-gated hierarchical policies — how we trained a 7-DOF arm to pour coffee in under 2,000 iterations.

Contact

Building in physical AI?

If you are working on data, research systems, or deployment infrastructure for physical intelligence, write to us.

Read the manifesto