“data, not compute, determines who wins”
The data layer for physical AGI.
We capture the whole-body, contact-rich, force-labeled movement that video can’t see.
Scroll
01 · The shift
AI left the screen.
For a decade, AI lived in text and pixels. The frontier has moved into the physical world — and the entire industry is converging on the same conclusion: the next defining AI companies will be built in atoms, not just bits.
a robotics “GPT-3 moment”
the constraint is no longer compute.
02 · The bottleneck
Large language models were trained on the open internet. Robotics has no such corpus. The data that teaches a machine how the physical world responds to action — force, contact, balance, grip — does not exist at scale.
Force and contact are invisible to video.
You cannot watch a video and feel how hard a hand grips before a bottle cap breaks its seal. That missing layer is the bottleneck to physical AGI.
03 · What we capture
The scarcest data in robotics.
Four streams. One clock. Each motion is captured as both what it looks like and what it feels like from the inside.
Whole-body pose
Per-joint position, orientation, velocity — 78 channels, 240 Hz.
Crown jewel
Fingertip contact force
The single scarcest modality in robotics — invisible to every camera.
Ground-reaction force
Sub-millisecond pressure across both soles. The densest signal in the stack.
Paired vision
RGB + depth on the same clock as every force sample.
04 · The stack
One stack. The full signal.
Haptic suit, instrumented gloves, predictive shoes, omnidirectional treadmill — synchronized on a shared clock.
Haptic Suit
100-motor haptic array, full-body. Captures whole-body pose and doubles as a contact-force surface.
Instrumented Gloves
Grasp and contact force at the fingertips — the single scarcest modality in robotics, invisible to every camera.
Predictive Shoes
IMU + pressure-sensor fusion. Ground-reaction force and gait — the densest-signal data in the stack.
Omnidirectional Treadmill
Unlimited locomotion capture in a fixed footprint. A structural data-cost advantage.
Predictive locomotion protected by US provisional patent.
05 · Flywheel
Data that compounds.
Every wearer-hour improves the model and is automatically scored by the model’s own prediction error. The data and the intelligence compound on the same loop.
Capture
Paired Vision + Force Data
Self-Supervised World Model
Prediction Error as Reward
Better Model
More Valuable Data
Data Flywheel
06 · Why HALO
Others capture slivers. HALO captures the signal.
Others
Slivers of the body.
- Grippers only
- Body pose without force
- Teleoperation — slow, expensive
- No fingertip contact data
- No ground-reaction data
HALO
The full signal stack.
- Whole-body pose
- Fingertip contact force
- Ground-reaction force + gait
- Paired vision on the same clock
- Synchronized at the source
07 · Why now
The capital and the science have arrived.
The data layer is still open.
$400M
Generalist AI Series B (~$2B valuation)
~$1B
AMI Labs seed (JEPA world models)
<1%
Frontier AI score on ARC-AGI-3 (humans: 100%)
2026
The robotics “GPT-3 moment”
How HALO captures force.
HALO Dynamics consists of a synchronized capture stack and a self-supervised world model. The stack records whole-body pose, fingertip contact force, ground-reaction force, and paired vision on a single shared clock. The model learns from the data without human labels — prediction error is the reward. Together they form an upstream layer every physical-AI team builds on.
AboutFor investors
The rare bridge.
Physical AI needs people who can build both the hardware that captures the world and the AI that learns from it. Almost no one can do both. HALO Dynamics was founded by a builder who shipped the full embodied stack — haptic suit, force-sensing gloves, predictive shoes, omnidirectional treadmill — and the autonomous AI systems to match. Predictive locomotion protected by US provisional patent.
VisionRead the long view
ImageNet for the body.
Frontier AI scores below 1% on ARC-AGI-3 while humans solve 100% of the same environments. The missing layer isn’t more text or more pixels — it’s embodied, contact-rich knowledge of how the physical world responds to action. That is exactly what HALO captures. The arc: HALO becomes the data and model substrate every physical-AI company builds on.
Learn more about what we do
The data foundry.
Buyer-grade, force-labeled human movement for the teams building physical AI.
DataThe capture stack.
Suit, gloves, shoes, treadmill — synchronized on one shared clock.
Hardware
Prediction-error as reward.
A JEPA-style world model trained on the data video can’t see. In development.
Research
Where it goes to work.
Humanoid labs, ergonomics, sports science, rehab, defense. One stream, five doors.
Product