Synchronized at the source
Every device shares one clock. Every force sample is anchored to a frame.
Data
The data foundry.
Licensed, buyer-grade embodied data for the teams building physical AI.
01 · What’s captured
The full signal stack.
- Whole-body posePer-joint position, orientation, velocity.
- Fingertip grasp / contact forcePer-finger normal + shear at the contact surface.
- Ground-reaction force + gaitSub-millisecond pressure across both soles.
- Paired vision + forceTime-aligned RGB / depth on the same clock.
- Full-body locomotionContinuous, unbounded via omnidirectional treadmill.
02 · Why it’s different
Not motion capture. Force capture.
Force-labeled, not just pose
What it looked like and what it felt like — captured together, every time.
Self-labeling
The human body provides the action labels. Physical contact provides dense force signals.
Hardware no one else has
Haptic suit, instrumented gloves, predictive shoes, omnidirectional treadmill — one rig.
03 · Data quality
Buyer-grade. Documented.
Buyer-grade means clean, synchronized, documented. Every dataset ships with a data card: capture schema, sensor specs, synchronization method, sample rate, and format.
- Object
- Glass bottle with screw cap, sealed
- Action
- Bilateral grasp → counter-rotate cap until seal breaks
- Devices
- Suit · Gloves · Shoes · Vision (RGB-D)
- Channels
- 162 (pose 78 / glove force 32 / shoe pressure 48 / vision 4)
- Sample rate
- 500 Hz force · 240 Hz pose · 60 Hz vision
- Frames
- 1,440 force / 144 vision · 2.4 s duration
- Sync method
- Hardware-triggered shared clock, ±0.5 ms
- Format
- Parquet (force/pose) · MP4 (vision) · JSON manifest
04 · Licensing
Two ways to access HALO data.
Dataset Licensing
Curated, contact-rich human-motion datasets for VLA and humanoid training.
Evaluation Environments
Paired data for benchmarking and policy evaluation.