phi9

phi9 is a physical AI company and research lab. It sells robot-learning data — egocentric video, full-body motion capture, and sim-ready demonstrations — while training toward physical intelligence methods, including a sub-5B VLA effort (phi9.zero). The north star is accelerating physical AI without collapsing into a generic robotics data vendor.

Purpose

Software agents are not enough when the job is to act under physics, contact, and embodiment. The bet is that world models for the physical world unlock more durable value than software-only copilots. Public framing lives at phi9.space/manifesto.

Operating model

The founding data loop:

1. Capture — synchronized multimodal motion and egocentric video
2. Multiply — relabeling and synthetic generation
3. Train — aligned pre-, mid-, and post-training streams
4. Evaluate — data and policies against benchmarks

GTM wedge: high-fidelity motion capture + synchronized video + structured task data, sequenced as pilot datasets → production datasets → custom capture → DaaS.

Research direction

Active VLA work favors native multimodal backbones with discrete action supervision and continuous control experts (implementation details stay in the private research tree). Related public research on this site that shares representation and verification discipline includes Measure Before You Predict and the RSI Lab portfolio map.

Public pointers

• Manifesto / lab: phi9.space
• Personal research index: shubhamattri.com/research