Two brains
for repeatable work
This page is for engineers, researchers and investors: what is inside a Heylper humanoid and how it learns. It builds on our research, where the brain was first proven on a real arm.
1.Two kinds of skill
Motor skills are a robot's muscle memory: grasp, place, probe, turn. They are generic, the same on every site. Procedural skill is how your expert does the whole job: which steps, in what order, what to check, what counts as done. It is specific to you, and it is what customers and auditors judge. A Heylper humanoid learns both on your job.
2.Inside a Heylper humanoid
Heylper's brain is two brains that work together, not a planner giving orders to the motors. The action brain moves the body. The procedural brain has three parts: a memory of your expert's procedure, a model that proposes, and a rule-based engine that decides.
The memory is the procedural memory graph (PMG): the steps, the goal of each one, the checks and the techniques, and why they matter. The model is NAP, our next-action prediction model. Every moment it reads the PMG and what the robot sees, proposes the next action toward the step's goal, and flags when the work starts to drift. The engine decides. It checks each proposal against the PMG and the world, answers the action brain when it asks, and when a step drifts, it stops the motion and replans the way back. NAP proposes, the engine decides: the same situation always gets the same decision, and every decision traces back to your expert's procedure. Both brains are trained together on your job, so the motions fit the procedure.
- The job"Service the control cabinet."
- PMGprocedural brain · remembersThe procedural memory graph: the expert's steps, goals, checks, techniques and recovery.
- NAPprocedural brain · proposesNext-action prediction: reads the PMG and the scene, proposes the next action every moment, and flags drift.
- Engineprocedural brain · decidesChecks each proposal against the PMG and the world, answers the action brain when it asks, and stops and replans when a step drifts.
- Action brainmotor policiesOur motor policies: one step in, motion out, adapted to where things are.
- The bodyopen hardwareOpen-source humanoid hardware, which keeps the robot affordable and easy to repair.
That is why it repeats. The common ways to give a robot a job leave the procedure unchecked while the robot works:
| Approach | Why the job drifts |
|---|---|
| One model for everything | A vision-language-action model turns each instruction into motion. It keeps no procedure, so the same job goes differently each run. |
| A planner on top | A language planner tells the motor skills what to do next. It plans from general knowledge, not your expert's, and sees a mistake only after the skill has run. |
| Copying demonstrations | Imitation learning repeats how the expert moved, not why. A new part or a shifted tool, and the robot drifts with nothing to catch it. |
| Heylper | Two brains work together. NAP proposes, the engine decides against your expert's procedure, so drift is stopped before the step goes wrong. |
3.What a procedure holds
A procedure is more than a list of steps. It is steps carried out in your place, with your things, alongside your people, and the PMG holds all three:
- Your stepsWhat to do, in what order, and what done looks like — isolate the power before opening the cabinet
- Your worldYour tools, where they are kept, your machines — the meter is in the second drawer
- Your peopleWhen to check, when to stop and ask, who to hand over to — ask the supervisor before restoring power
The steps are captured at four levels of detail, shown here for the job in the concept films:
- L1 · TaskThe goal — "Service the control cabinet"
- L2 · StepsThe plan — isolate the power → open → test → service → restore
- L3 · ActionsThe doing — switch off the breaker, lock and tag it, open the door
- L4 · TechniquesThe quality — prove the meter on a live source before trusting a dead reading
4.How it learns: the Expert Hat
The Expert Hat is a role, not a device: whatever watches, listens and guides on the expert's behalf. On a Heylper humanoid it is the robot itself, so the robot is an apprentice. It learns the procedure, not the motions, in three beats:
- LearnIt watches the expert demonstrate, asks why, and captures the procedure: the order, the checks, the technique.
- RefineIt guides the expert's colleagues through the job. Where it is unsure, the expert corrects it, so the gaps close before it works alone.
- GuideIt does the job itself, and every run is checkable against the procedure.
Each kind of skill has its own pipeline, and Heylper runs both:
| Motor skills | Procedural skills | |
|---|---|---|
| Learned from | teleoperation demonstrations | expert demonstrations |
| Refined in | simulation | guidance sessions with people |
| Deployed as | a policy on the robot | a procedure the robot follows |
| Made by | Heylper | Heylper |
The two pipelines feed each other: the expert data gathered for a procedure also refines the motor policies, so they act faster on your job. In the first pipeline the simulator is software. In the second, the simulator is your own people — and the sessions that refine the skill are the same sessions that train them.
Back to home