// PRE-SEED · POLICY TRAINING · DEXTEROUS HARDWARE · US-BUILT

Robots that learn the task. Hands built to do it.

Our pipeline trains manipulation policies for industrial robots automatically, with no hand-tuned rewards and every result measured on conditions it never trained on. It is heading onto real hardware now. Alongside it, we build the Dex Hand, a 6-DOF dexterous end-effector designed alongside that software.

See the software The Dex Hand
NVIDIA Inception Program member

01 — What we do

Two products, built together and moving fast.

The software trains the skills. The hardware gives those skills somewhere to go next. The software works on the grippers already in service, and real hardware is the next step for it.

Phase 1 ready now

Policy training pipeline

A task in plain English and a model of your robot in; a trained, verified policy out. No one hand-tunes a reward.

  • Now: trained policies per task, verified in simulation, for parallel-jaw grippers
  • Ranked on held-out success, never on reward
  • Delivered with its success rate per condition, and footage
  • Next: the pipeline itself, licensed to automation teams

We want the pipeline in companies' hands as soon as possible. What stands between here and there is proof on real hardware. If you run a cell, that is where you come in.

Offer a real-hardware cell How it works
Phase 2 in manufacturing

Dex Hand

A 6-DOF dexterous hand for industrial and humanoid applications, built to run policies from the same pipeline.

  • Precision linear actuators, onboard custom MCU
  • Joint position, fingertip force, wrist force/torque
  • URDF model, ROS 2 packages, trained task policies
  • US-designed and assembled
Specs & roadmap

02 — The opportunity

Trainable skills and dexterous hands turn 4.66 million installed robots into general-purpose machines.

Arm motion is solved. So is vision. Both are deployed at scale, and what multiplies them is a robot that can be taught a new task quickly and a hand that can carry it out. Today every part, orientation and line change needs an expert to reprogram it, and the clamp on the end does only what it was fixtured for.

That is why assembly, kitting and machine tending are still done by hand at most plants. We build both pieces: skills trained without an expert in the loop, and an end-effector dexterous enough to use them.

Dex Hand Phase 2 assembly, open pose

03 — Software

Describe the task. Get a verified policy.

Following NVIDIA’s Eureka method, a language model writes candidate rewards, our pipeline trains them in simulation, and every candidate is graded on its measured success under conditions it never trained on. The months of expert reward tuning that robot learning usually needs, repeated for every task and every part, happen inside the pipeline instead.

Failures are catalogued and checked on every future run, so the system gets better with each task. Adding a new task becomes a simulation run, not a hiring plan.

How the software works
Dex Hand mounted on 6-axis arm, cell context
95% Held-out pick-and-place, parallel-jaw gripper (sim)
0 Reward weights tuned by hand
~1 day To bring up a new parallel gripper

// Simulation results in NVIDIA Isaac Lab, deterministic policy, on approach distances it never trained at. Real-hardware validation is the next milestone.

04 — Dex Hand

A dexterous hand, designed alongside the policies that run it.

The Dex Hand comes up through the same training pipeline as every gripper we support, so it arrives with policies for the target task rather than as a bare mechanism the customer must program from scratch.

Not just a URDF and a driver package, but trained policies for the target task. The declarations our pipeline needed for a commercial gripper are the same ones a multi-finger hand needs, so the work on one carries straight over to the other.

Dex Hand fingertip detail, sensor array
Actuation
6 active DOF, 5 passive · precision linear actuators
Sensing
Joint position, fingertip force, 6-axis wrist force/torque
Policy pipeline
LLM-generated rewards, Eureka method · RL training in NVIDIA Isaac Lab · sim-to-real transfer
Control
Onboard custom MCU · self-contained end-effector
Integration
URDF model · ROS 2 packages · trained task policies
Origin
US-designed and assembled · auditable supply chain

05 — Market

The market is already installed.

4.66M Robots in operation
542K Installed in 2024
700K+ Annual, forecast 2028

Every one of those robots needs its tasks taught. The software is built for that installed base, on the grippers already in service. The hand reaches the tasks those grippers cannot.

// Source: IFR World Robotics 2025 · Robot end-effector market $5.54B (2025) → $10.58B (2030), Mordor Intelligence

06 — Roadmap

Two tracks, moving in parallel.

Software

  • Ready now

    Phase 1 — Trained policies per task

    The pipeline trains and verifies a policy for a given gripper and task, in simulation. Pick-and-place solved end to end at 95% held-out with no hand tuning; an industry-standard adaptive gripper is going through the same stages now.

  • Critical next step

    Phase 2 — Real-hardware validation

    Real cells, real success rates reported next to the simulated ones.

    This is the step that lets us sell the pipeline. We are pursuing it now, and we need hardware: a gripper, an arm and a real task. If you can offer a cell, talk to us.

  • As soon as Phase 2 proves it

    Phase 3 — The pipeline, licensed to companies

    Automation teams run the pipeline themselves on their own robots and tasks. More end-effectors, more task families.

Hardware

  • Complete

    Phase 1 — Mechanism

    Fully 3D-printed and servo-driven. Proved out the kinematics.

  • Design complete

    Phase 2 — Software test platform

    Aluminum frame, linear actuators, custom MCU, full sensor suite. Moving into manufacturing and assembly, built to prove the software against real parts.

  • Next

    Phase 3 — Production

    The hand that ships, designed for volume manufacturing from the start.

We are raising a pre-seed round.It funds both tracks: real-hardware validation of the pipeline, and Phase 2 hands into production cells.

Investor enquiry →

Follow along as we build.Results, failures and what they taught the system, on the blog.

Read the blog →

07 — Team

The founders.

Kameron Jones

Kameron Jones

Co-Founder

Lead Engineer — mechanism, tolerancing, manufacturing

Milo Kucia

Milo Kucia

Co-Founder

Software Lead — reinforcement learning, control, firmware

08 — Contact

Start a conversation.

We are raising a pre-seed round to take the pipeline onto real hardware and put Phase 2 hands into production cells. If you are an investor, an integrator, or an automation team with a cell that could be our first real-hardware test, we would like to hear from you.

Offer a real-hardware cell Investor enquiry