Same task, same way, every time.
A robot that can do the job still does it differently each time.
With Heylper, it does it your way in every session.
1.Capable is not the same as repeatable
Today's robot foundation models, the VLAs, take one instruction at a time and pick a reasonable way to carry it out. Run them again and they pick another reasonable way.
For a new task that is a feature. For work that already has a procedure, it is a failure. A lab, a factory floor or a service bench has a standard operating procedure: an order that must hold, checks that must be made, things that must never be done. A run that ignores it has failed, even when the job looks done. The procedure is not the robot's to choose.
2.On a real arm: loading a centrifuge
A centrifuge rotor has to be balanced: two tubes go into opposite slots, and the lab's procedure says which ones, slots 1 and 5. We gave an SO-101 arm the same task in ten sessions: five on its own robot skills, and five with Heylper holding it to the procedure.
SO-101 arm · load two tubes into opposite slots of a centrifuge rotor · five sessions each, shown at 10× speed · watch on YouTube
| Session | Without Heylper | With Heylper |
|---|---|---|
| 1 | Slots 2 and 4 · not balanced | Slots 1 and 5 · balanced |
| 2 | Slots 1 and 4 · not balanced | Slots 1 and 5 · balanced |
| 3 | Slot 5 and the centre hole · not balanced | Slots 1 and 5 · balanced |
| 4 | Slots 1 and 5 · balanced | Slots 1 and 5 · balanced |
| 5 | Slots 2 and 4 · not balanced | Slots 1 and 5 · balanced |
| Balanced | 1 of 5 | 5 of 5 |
On its own, the arm found a different answer almost every time, and once put a tube in the centre hole. With Heylper it did the same thing in every session: tubes from the same rack, into slots 1 and 5, as the procedure says.
3.Why it works: the same decisions, not the same motions
Heylper does not make the arm move identically every time. Conditions vary, and adapting to them is what the robot's motor skills are for. Heylper makes the decisions identical: which step comes next, where, and what counts as done.
- Heylperthe expert brainRemembers the procedure as a graph of steps, each with a check. The same situation in gives the same plan out.
- Modelseyes, not judgesAnswer questions about the world, such as "is a tube in slot 5?", and never decide the procedure.
- Motor skillsthe robot's ownA VLA or ROS tools. One instruction in, motion out.
Because the plan comes from the procedure, any difference between two runs points to a specific fact that differed: the robot saw something else, or a step did not do what it should. That makes every run checkable afterwards, by you or by an auditor.
4.The research behind it
The design is set out in our working paper, Executable Procedural Memory Graphs: Remembering Procedures, Not Policies (September 2026). Here is how its terms map onto the Expert Hat.
| In the paper | In the Expert Hat |
|---|---|
| Procedural Memory Graph | The expert's procedural skill: the steps, their order, the checks and what never to do. Captured in Learn, sharpened in Refine. |
| Reasoning engine | The expert brain: it gives the robot the next step, and does not move on until the last one is done. |
| System 1 | The robot's own motor skills: a VLA, or tools on ROS. |
| Grounded verdict | The check after every step, made independently of the robot reporting "done". |
| Provenance and open questions | Where each part of the skill came from, and what is still uncertain: the questions it takes back to the expert in Refine. |
| Divergence classes | When a run goes wrong, Heylper says why: a wrong rule in the procedure, a misread of the scene, or a motion that failed. |
5.CoRL 2026: selected for the demo lineup
Our submission, Procedural memory guided robotics for repeatability and safety, was selected for the demo lineup of the Agentic Robotics Workshop at CoRL 2026 in Austin. The workshop asks the question Heylper answers: where the line should fall between a robot's reasoning and its motor skills.
A simulated arm placing blocks. Left: the robot's motor skills acting in simulation. Right: the working memory graph and the planner walking the procedure, step by step · watch on YouTube
6.What this shows, and what it doesn't yet
- ShownOne procedure held across sessions on a real arm, 5 of 5 against 1 of 5, and the planner walking a procedure graph live in simulation.
- Not yetResults at scale. Five sessions each is a demonstration, not a benchmark, and the paper's own result is preliminary.
- NextA fixed-scene repeatability benchmark, with and without Heylper, on the open-source reBot B601 arm: our first product.
7.From here to your robot
The Expert Hat comes first to the open-source reBot B601 arm, then to humanoid workers. The procedure it learns from your expert is the same on both.
Back to home