Robot intelligence built to steer through the real world.
The best robot architectures only work with the right data behind them. We close that gap by pairing elite data with frontier models, finally moving AI from digital to physical.
We capture both sides of every manipulation.
Humanoids will reshape physical labour. The bottleneck is not mechatronics, it is data. Robot foundation models learn from teleoperation and video, and neither records what the hand actually felt. We capture what they miss, then test and scale our models on JEPA architecture.
At a glance
- Signals
- Tactile, motion, vision
- Alignment
- At capture
- Coverage
- Full hand surface
- Correspondence
- 1:1 human to robot
Demonstration
Human · glove
Deployment
Robot · skin
Notes from the lab.
Research notes, design choices, and field reports from our work on capturing manipulation data at scale.
The data gap is not our hypothesis.
It is openly acknowledged by the best-funded players in the field. Every one of them has published the same conclusion: robots need manipulation data at a scale that teleoperation and video cannot reach, and video does not record force at all.
“Generalist robotics have arrived, powered by advances in mechatronics and robot AI foundation models. But a key bottleneck remains: robots need vast training data for skills like assembly and manufacturing tasks. Traditional robot foundation models require extensive manual demonstrations for every new task and environment, which is not scalable.”
NVIDIA
2025
Enhance Robot Learning
“Without tactile sensing, robots depend on video to interact with their environment. With video alone you don't know you've touched something until well after the collision has physically caused the object to move. This reduces work efficiency and can require numerous attempts, grasping and re-grasping the same object for a secure hold. Touch solves this.”
Sanctuary AI
2025
General Purpose Robots
“Videos do not show the underlying forces, torques, or tactile feedback… A human hand, a 7-DOF industrial arm, and a quadruped all have vastly different morphologies. Mapping a human's 'grasp' to a robot's 'actuation' is a massive translation problem.”
Skild AI
2025
Learning by Watching Human Videos
“…pre-training on 20,854 hours of human egocentric video spanning 20+ task categories, from manufacturing and retail to healthcare and home environments. This is a significant step up from the few thousand hours of robot teleoperation data… Training on sensorized human video (ego cameras, wrist cameras, hand tracking) gives the model rich manipulation priors without requiring every behavior to be demonstrated on a physical robot first.”
Hugging Face
2025
Open Reasoning VLA Model for Humanoid Robots
“Collecting real data at scale with task-centric ground truth labels, like contact forces and slip, is a challenge further compounded by sensors of various form factor differing in aspects like lighting and gel markings.”
Meta
2025
Self-supervised Touch Representations
Our answer
The best-funded labs in robotics are already paying for human video because of the performance gains it unlocks. Video gives a flat, partial view of the hand and none of the forces. We capture the reaction as well as the intent: tactile, motion and vision, aligned at the moment of capture.
SoftShell
Let's talk.
We work with a small number of humanoid programs, prosthetic manufacturers, research labs and serious investors at any time. Tell us briefly who you are and how you'd like to engage.
Loading form…