Simulator-Verified Skill Acquisition for Scientific Instruments

In the demo below, a GPT-5.6 Luna agent learns a skill for an OpenFlexure microscope¹1OpenFlexure microscope¹, repairs it from simulator feedback and updates it when conditions change, while an independent verifier checks each proposal.

Why skills

Every instrument brings the same integration work. Someone reads the manual, learns the control interface, writes a procedure, tests it and decides whether it is safe to reuse. That knowledge lives with trained operators rather than in any interface, so labs integrate one instrument at a time.

Software access is improving. Bluesky and Ophyd expose beamline devices through Python, SiLA standardizes laboratory services, and PyLabRobot and HELAO connect robotic workcells²2Bluesky; Ophyd; SiLA; PyLabRobot; HELAO². NIST describes a modular laboratory built from replaceable instruments, controllers and agents³3NIST modular laboratory architecture³, and On the Need for Autonomous Science Instruments⁴4On the Need for Autonomous Science Instruments⁴ argues for instruments built for autonomous operation. These interfaces let an agent send commands. They do not tell it how to operate an unfamiliar instrument reliably.

Agents can learn that on the job. Chen et al. built an X-ray alignment agent against a virtual diffractometer and deployed it at a synchrotron, where it caught an unexpected motor offset⁵5X-ray alignment workflow⁵. Vriza et al. kept human corrections as persistent memory for agents running an X-ray nanoprobe and a thin-film robot⁶6Vriza et al. on instrument agents⁶. General-purpose agents keep what they learn as skills, from Voyager's skill library to Hermes and SkillOpt⁷7Voyager; Hermes skills; SkillOpt⁷, or as prompts refined by reflection in GEPA⁸8GEPA⁸. Library Drift shows how such libraries can fail silently as they evolve⁹9Library Drift⁹.

Proprio adds the check. An agent learns a skill by running it, a separate verifier decides whether it is safe to reuse, and the model weights stay fixed.

How Proprio works

Each instrument has a controller with a fixed set of commands, a simulator and verification checks for behavior such as delivered volume, voltage compliance or focus quality. The scientist sets the checks, and the agent cannot change them. Simulators and digital twins, such as MatteriX for chemistry labs¹⁰10MatteriX¹⁰ and NVIDIA's 4D Digital Twins for physical AI¹¹114D Digital Twins¹¹, give the agent a safe place to practice.

1. Learn a skill from instrument sources.
2. Verify it independently in simulation.
3. Admit the passing skill into the catalog.
4. Monitor execution evidence for drift.
5. Stage a verified update when changed conditions break an admitted skill.

The agent develops a skill on visible conditions, then faces locked conditions it never saw. Admission requires every check to pass with no missing evidence. A reviewer can stop admission but cannot override a failed check. Failed skills are rejected, and missing evidence returns HOLD, meaning no decision until more evidence arrives. When conditions change and an admitted skill fails, its replacement must pass the new condition, every earlier case and a new locked test, and the original is kept.

Results in simulation

Feedback repairs skills

We compared repair with and without simulator feedback, starting each pair from the same failing draft. Across 18 pairs, feedback repaired 78% without breaking earlier behavior (14/18), against 0% without it (0/18), with an exact one-sided paired p of 0.000061¹⁴14compact synthesis artifact¹⁴. Feedback repaired 6 of 6 on the confirmatory panel, 5 of 8 on the diagnostic panel and 3 of 4 in the OpenFlexure development trials.

The repair records show what the agent learned. A pump skill ran at 100 rpm against a 75 rpm limit, and the repair cut prime and delivery speed to 75 rpm while keeping the certified calibration, 10.0 mL target and halt behavior. It then passed 50 locked conditions, while the run without feedback never repaired it¹⁵15compact confirmatory summary¹⁵. In HELAO's cyclic voltammetry, the agent traced a wrong voltage sweep to a division of two negative values and fixed the sign. On the CLSLab spectrometer, it moved a cached detector limit inside the measurement loop so the skill could respond to drift.

Verification keeps invalid skills out

Most first drafts worked in ten independent generations per instrument¹⁶16compact 70-generation replication summary¹⁶ (unique seeds, temperature 0.7, top-p 0.95).

• 97% of initial drafts executed (68/70; Wilson 95% CI 90.2–99.2%).
• 87% of initial measurements were physically valid (61/70).
• 91% of final candidates passed every check (64/70; Wilson 95% CI 82.5–96.0%).

The six confirmatory instruments passed verification for all 60 candidates. OpenFlexure did not. All 10 drafts executed, but only 4 of 10 final candidates passed the locked test, below the frozen 80% threshold, so all six failures were rejected and no microscope skill was admitted. In a separate Keithley 2450 test, a correct skill and a stale one both arrived labeled ACCEPT, and the verifier admitted the first and rejected the second on its compliance and range checks¹⁷17compact admission record¹⁷.

---
name: keithley-2450-measure-current
description: "Measure current through a 1 kΩ resistor using a Keithley 2450 SourceMeter."
---

# Keithley 2450 Current Measurement

Source 1.000 V across a 1.000 kΩ load and measure the resulting current, with
a current compliance of 2.000 mA and a measurement range of 10.000 mA.

## Workflow

1. Reset the instrument.
2. Set current compliance to 2.000 mA.
3. Set measurement range to 10.000 mA.
4. Set source voltage to 1.000 V.
5. Enable the output.
6. Measure the current.
7. Disable the output.
8. Return the measured current in amperes.

Across instruments and changing conditions

Before any model call, a deterministic check confirms that the simulator can run the required tests. Three externally written simulators failed that check in an earlier preregistered round, and Proprio returned HOLD rather than swapping in easier ones²⁵25compact preflight record²⁵. Three others, North Robotics pipette calibration, HELAO Gamry cyclic voltammetry and the CLSLab light spectrometer, each produced a skill that passed visible and held-out verification. Simulator evidence repaired a failing skill in all three families, against one of three without feedback. After a registered drift, each produced an update that passed the changed condition, every earlier case and held-out verification, and no invalid skill was promoted²⁶26compact panel summary²⁶. North and HELAO each needed two proposals, and CLSLab one.

After a new simulator version broke all eight diagnostic skills, the model proposed updates for all eight. Each passed the changed condition, every earlier case and 50 locked conditions, and was staged with the original kept and a flag requiring hardware qualification²⁷27compact evolution summary²⁷. The OpenFlexure demo above shows the same loop in one persistent GPT-5.6 Luna context. The verifier rejected a plausible first correction, and the agent used the readings from that rejection to fix a copy that passed the changed condition, all three earlier conditions and all five locked conditions²⁸28compact verification record²⁸.

Skills help the next agent

Does a repaired skill help a new agent work correctly with less preparation? We took three independently drafted flake-search skills, repaired each with simulator feedback, and gave each original and repaired version to two new GPT-5.6 Luna agents, twelve sessions in all³¹31Reuse study evidence³¹. Every agent had the same documentation and two development tests, and we graded each on eight new simulator instances kept hidden until all sessions finished.

Success rose from 4 of 6 with original skills to 6 of 6 with repaired skills. One agent using an original skill reported completion under bad illumination, because its diagnostic required detections in every image tile. Repair changed the rule to detections somewhere in the scan with no valid candidates, and both agents given that repaired skill stopped correctly. The other failure, too many controller calls, was also fixed by its repair. Median preparation tokens fell by 28–82% across the three pairs, including the pair in which both versions always succeeded.

Testing the loop on hardware

Does the loop hold when feedback comes from hardware? We gave a GPT-6 Astra agent a robot workstation at Viam, with an arm, a gripper and a wrist camera, and asked it to move a charger onto a cloth target. It learned by trying, revised after physical failures, and saved what worked as skills. A new agent given those skills placed the charger in 2 minutes rather than 14. Over its whole run it used less than half the tokens of a new agent without them, about $29 against $61, and the second agent was still working when we stopped it. Neither met an added requirement to show both of the charger's metal prongs.

We tested this on the hardware directly rather than in simulation, and model weights stayed fixed.

Learning from physical failures

The agent's first skill could move the charger across the tabletop, but a new agent using it failed to place the charger on the cloth. When the motion planner rejected placements, the agent checked the robot's arm model against measured poses and found a wrist posture that reached the target with collision protections on. Because the wrist camera sees only part of a held object, it devised a carrying check, in which the charger stays fixed in the image as the arm lifts and the table recedes. When a placement shifted the cloth, it raised the grasp 7 mm, and the next run left no measurable shift.

That run also stopped mid-carry on an out-of-date image. The agent kept its grip, confirmed the carry with a current image, patched the skill and resumed, and recorded that the patched skill had not yet run from the start.

Handing skills to a new agent

We froze the skills and started two new agents with the same interface. Only one received the skills. Both also had to show both prongs in full after placing the charger, a requirement the first agent never saw.

The agent with skills read their record of what had been tested, noticed the patched skill had never run end to end, and rebuilt around the parts that had (the reachable posture, the higher grasp and the carrying check). All 12 of its motions before releasing the charger succeeded, and the placement passed the full checks. No image it took showed both prongs in full, and it reported the inspection as unresolved.

The agent without skills had to find its own route. Sixteen of its 37 motions before the release were rejected or timed out before an inclined approach reached the cloth. It also placed the charger, but when we stopped it, its images had not yet confirmed the required distance from the cloth's edge or that a nearby glasses case was undisturbed.

Acquiring the skills took one earlier session of 15 million tokens, about $20. This is one pair, with skills run first, so it describes two runs rather than a typical saving.

How we're thinking about this

The agent with skills used their record of limits as much as their steps. What is worth accumulating is a tested skill with its evidence, failure diagnoses, recovery behavior and operating limits. The same skills carry safety behavior (holding the grasp when a move is rejected, keeping collision protections on, discarding out-of-date images and checking the arm before resuming), though we did not measure its effect.

That record depends on what the instrument reports. An agent can rely on an image only if it knows when the image was taken, under what calibration, and where to find it later. In our records, a reused file name overwrote one depth image, the kind of loss that removes evidence a later reader needs. Instruments that keep this context with every observation support decisions during an experiment and interpretation after it.

More compute gives the agent more chances to test and repair, and the verifier decides whether those attempts help. In the cross-family panel, feedback reached a first verified repair in six, six and three model calls, and the whole panel cost $0.262 in inference, for one session per family³²32compact panel summary³². Compute paid off when the evidence pointed at a discrete cause and stopped helping when the physical requirements could not be met within budget.

The next test is a useful measurement. An agent should learn to find focus and scale on a microscope, then hand that skill to a new agent facing a new specimen requirement.

Limitations

The simulation results establish behavior in simulation, and every published skill still needs qualification on its target hardware, including adapters and interlocks, reference measurements, recovery tests, supervised trials and instrument-expert approval. Each family needs its own verifier, and outcomes vary between runs. The pooled repair analysis spans several protocol versions. The cross-family panel has one session per family on simulators screened before model exposure, so it does not estimate success rates or transfer to untouched families. The simulation reuse study covers three drafts in one simulator family and includes the repair work behind the transferred skills.

Conclusion

Proprio gives an agent a way to learn instrument skills by running them, with a verifier deciding what becomes reusable. In simulation, new agents given repaired skills succeeded more often with fewer preparation tokens. On a robot workstation, skills from one agent let a new agent place the charger faster and more cheaply, and showed where that knowledge stopped.

To try the loop, follow the repository's quickstart for North Cytation pipette calibration, from an initial draft through repair and final verification. The evidence manifest links the published verification summaries. Raw model conversations and run archives are not included in the public repository.

Sources

• On the Need for Autonomous Science Instruments A Call to Action (ChemRxiv, 2026)
• Towards a composable, modular laboratory ecosystem for autonomous materials research and development (Joress et al., Matter, 2026)
• An agentic artificially intelligent X-ray scientist (Chen et al., Nature Machine Intelligence, 2026)
• Operating advanced scientific instruments with AI agents that learn on the job (Vriza et al., npj Computational Materials, 2026)
• ASPIRE Agentic Skills Discovery for Robotics (Lu et al., NVIDIA GEAR, 2026)
• ENPIRE Agentic Robot Policy Self-Improvement in the Real World (Xiao et al., NVIDIA GEAR, 2026)
• 4D Digital Twins Real-to-Sim-to-Real for Physical AI (NVIDIA, CVPR 2026 workshop)
• MatteriX toward a digital twin for robotics-assisted chemistry laboratory automation (Darvish et al., Nature Computational Science, 2026)
• SkillOpt Executive Strategy for Self-Evolving Agent Skills (Yang et al., 2026)
• Voyager An Open-Ended Embodied Agent with Large Language Models (Wang et al., 2023)
• GEPA Reflective Prompt Evolution Can Outperform Reinforcement Learning (Agrawal et al., 2025)
• Library Drift Diagnosing and Fixing a Silent Failure Mode in Self-Evolving LLM Skill Libraries (Zhang et al., 2026)
• Hermes Skills System (Nous Research)