EightDim trains models of the physical world across all eight dimensions — six spatial, time, and the one every other stack leaves out: tactile. Touch data captured by a tendon-hydraulic hand and devices cheap enough to put in a million hands.
We are training the physical world with the tactile dimension added. Six spatial dimensions, time, and touch — every layer of the stack is built to capture, simulate and learn all eight, so they reinforce each other.
A differentiable soft-body engine that models tendon elasticity, fluid pressure and contact at sub-millimeter fidelity — 10,000× faster than real time.
Foundation models that don't just imitate grasps — they generate novel manipulation strategies for objects and tasks never seen before.
A sub-$200 wearable capture device that turns any human hand into a training sensor. Cheap enough to deploy by the thousand.
A tendon-hydraulic hybrid hand: the compliance and routing of biology, the raw force density of hydraulics. Human-grade dexterity, machine-grade endurance.
Evolution spent 60 million years proving that tendon routing is the right architecture for dexterity. We kept the architecture — and swapped the muscles for hydraulic power.
Bio-inspired antagonistic tendon pairs route force through anatomically accurate pulleys, delivering compliant, back-drivable fingertip control.
Micro-hydraulic actuators deliver force density no electric motor can match — silent, smooth, and thermally unbreakable under continuous load.
Manipulation data is the bottleneck of robot learning — so we made collecting it radically cheap. Our wearable capture rig costs less than a pair of headphones, streams full hand kinematics, contact forces and first-person video, and works anywhere: kitchens, factories, workshops.
Low-cost wearables record human manipulation in the wild, at fleet scale.
Physics simulation multiplies every episode across randomized worlds.
Manipulation models learn to synthesize new strategies, not just replay old ones.
Policies run on the hand; real-world outcomes flow back into the dataset.
Most simulators treat a hand as rigid links and torque motors. Ours models what actually matters: tendon stretch, pulley friction, hydraulic pressure waves, soft fingertip deformation and multi-point frictional contact — all differentiable, all GPU-native.
We're partnering with select robotics teams, labs and manufacturers for early access to the EightDim hand and 8-dimensional data platform.