Reinforcement learning for the factory floor

Precision robotics that absorbs variability.

Real parts shift, flex, and slip. We give industrial robots small learned policies that correct for it in real time, on top of the controllers you already trust.

Start a pilot How it works

Policy size: ~36Kparams
Runs on: CPUno GPU at the cell
Control rate: 50Hz

pose_error / trial classical + learned correction

in tolerance — to —

Pick & placeInsertionAssemblyMachine tendingKittingDeburring

01 — The problem

Automation works until the world varies.

Small errors, big failures

A few millimeters of pose error or an unexpected contact turns a reliable motion program into a missed grasp, a jam, or a scrapped part.

Yield ramps take months

New stations often launch well short of target yield. Closing the gap means endless edge-case tuning, fixtures, and re-teaching.

Big models aren't the answer

General-purpose robot foundation models are slow, costly, and hard to certify. A welding cell doesn't need a humanoid's brain.

02 — Our approach

Classical control for what you can model. Learning for what you can't.

In manufacturing you already know the part, its CAD, the robot, and the task. We keep your deterministic controller in charge and train a compact residual policy that makes small, bounded corrections from force, torque, and robot state.

Bounded by design. Corrections stay inside limits you set, so the cell stays predictable.
No pixels required. Policies learn from contact feedback, which narrows the sim-to-real gap.
Trained in simulation. Thousands of randomized variations of your cell, in minutes on a single GPU.

new

Learned residual

Tiny policy · Δx, Δy, Δyaw, timing

yours

Nominal controller

PLC / robot program · trajectories, IK

Robust motion

Recovers from shifted parts and uncertain contact

03 — From cell to policy

A pilot in weeks, not quarters.

01
Model the cell

We import your robot, gripper, part CAD, and existing motion program into a physics simulator.
02
Randomize reality

Pose error, friction, mass, sensor noise, and latency are varied across thousands of parallel runs.
03
Train & verify

A compact policy is trained and scored on paired trials against your baseline, failure by failure.
04
Deploy at the edge

The policy ships as a small artifact that runs beside your controller on standard industrial hardware.

04 — Platform

Two ways to work with us.

Precision Layer

Solutions

We build, validate, and deploy learned corrections for your hardest stations. You get higher first-pass yield without redesigning the cell.

Grasp recovery under pose uncertainty
Contact-rich insertion and assembly
Paired-trial reports you can take to the plant manager

RL Toolkit

Tools

Easy reinforcement learning for robotics teams. Go from a robot program to a trained, tested residual policy without a research lab.

Cell templates for common arms and grippers
Domain randomization and GPU-parallel training built in
One-command export to an edge-ready policy

$ solved init --robot ur5e --gripper 2f85
$ solved randomize pose friction mass latency
$ solved train --envs 1024
$ solved eval --paired --baseline nominal
✓ policy.onnx  36K params  ready

05 — Who it's for

Built for the people who ship automation.

System integrators

Add robot learning to your offering without hiring a research team. Win the stations your competitors walk away from.

Manufacturers

Raise yield on existing lines and shorten bring-up on new ones, with the equipment you already own.

Primes & OEMs

Handle high-mix, high-value parts where every drop or misalignment is expensive, and where determinism matters.

Have a station that won't hit yield?

Tell us about the part, the robot, and where it fails. We'll tell you whether a learned correction can fix it.

Get in touch