Data

Building the highest-fidelity sensorimotor dataset for physical intelligence.

Multi-modal, real-time synced, and exportable straight into pretraining, policy, and world-model pipelines. Phi9 captures the full shape of human behavior — manipulative and locomotive, body and hand, intent and failure — at the fidelity training and evaluation actually need.

  • Tactile force feedback
  • Motion
  • Vision
  • Depth
  • Wrist POV
  • IMU

  • 12+

    Modalities per capture

  • 25+

    Environment types

  • 3

    Training pipelines

Capabilities

What every capture carries.

Twelve capabilities ship with every phi9 capture — not a menu to pick from. They are recorded together, timed to the millisecond, and structured so each one can feed a pretraining, policy, or world-model pipeline without the others being rebuilt.

Highest-fidelity motion

120+ Hz full-body pose with body and hand trajectories. Calibrated per session, minimal drift.

Real-time multi-modal sync

Every modality — motion, video, depth, IMU, tactile — aligned to within a millisecond at capture.

Manipulative + locomotive

Fine-motor tool use and whole-body movement, captured on the same rig without a setup swap.

Parent + subtask labels

Hierarchical demonstration structure — parent task, subtasks, intent, success, and failure.

Pipeline-ready exports

Drop straight into pretraining, policy training, or world-model pipelines. Format-native, simulation-ready.

Multi-view video

Ego-centric and wrist POV streams together — first-person context plus close-object detail.

Depth

Stereo and time-of-flight streams where geometry carries information the RGB does not.

IMU

Wrist and body inertial signals. Context, redundancy, and sub-motion detail the cameras miss.

Tactile

Glove-level force feedback for contact, grasp, and manipulation states.

Scene metadata

Object state, lighting, layout, and calibration quality — the full context around every capture.

Retargeting-ready

Skeletal maps export cleanly to humanoids, robot arms, and simulated agents.

Evaluation traces

Success cases, failure modes, and transfer results across setups and environments.

Outputs

Three training pipelines, one capture.

The same demonstration feeds pretraining data for foundation work, policy training data for robot skills, and world-model data for simulation and planning.

  • 01 Pretraining data
  • 02 Policy training data
  • 03 World-model data
Custom capture

Request data your research actually needs.

Your research should not stall while you build a data pipeline. Phi9 captures custom, private, proprietary datasets — scoped to your task, your environment, and your evaluation criteria. We run the rig, synchronize the modalities, label the demonstrations, and hand back pipeline-ready data.

Email research@phi9.space See the lab

Contact

Building data systems for physical AI?

If you are working on collection, structured datasets, or research infrastructure for embodied systems, write to us.

See the lab