Dex Hand

Arm motion is a solved problem. So is vision. What's still holding automation back is the part bolted to the wrist — a clamp built for exactly one part in exactly one orientation, that needs a new fixture and a re-tool every time the line changes.

That's why assembly, kitting, and machine tending are still done by hand at most plants. Dexterity is the missing piece, and it's what finally lets automation reach the jobs it's never been able to touch.

Getting a hand to reliably grasp real parts is mostly a reward-shaping problem — the unglamorous, expert-hours work of hand-writing and re-tuning reward terms for every new task, on every new part. That's the piece of the year-long integration slog that's hardest to shortcut.

We build on NVIDIA's Eureka method: an LLM generates candidate reward functions, we train and evaluate them in simulation, and the loop rewrites itself toward better grasps instead of a person iterating by hand. Adding a new task becomes a simulation run, not a new hire. Paired with hardware that's US-designed and assembled on an auditable supply chain, that's what actually gets a hand from bench to line fast.

Phase 1 answered the mechanical question — can a low-cost hand actually replicate human-level dexterity? Fully 3D-printed and servo-driven, it proved out the kinematics before we committed real budget to it.

Phase 2 answers the harder question: can the reward-synthesis pipeline actually get a real hand to grasp real parts, outside simulation? The design — aluminum frame, precision linear actuators, a custom onboard MCU, full sensor suite — is finalized in SolidWorks and heading into manufacturing and assembly. It exists to run our software against real tasks, not to ship as a product.

Phase 3 is the production hand: whatever Phase 2 proves out about the mechanism and the pipeline, built into hardware designed for volume manufacturing from the start.