---
title: "A robot builds a laser, then realigns it when you knock it out"
locale: "en"
url: "https://irz.fr/en/articles/robotic-optics-protocol-en"
markdown_url: "https://irz.fr/en/articles/robotic-optics-protocol-en.md"
category: "tech"
tags: ["robotics", "optics", "self-driving laboratory", "MIT", "scientific automation"]
published_at: "2026-09-23T09:08:00.000Z"
author: "Arthur Lacoste"
translation: "https://irz.fr/fr/articles/robotic-optics-protocol-fr.md"
---

# A robot builds a laser, then realigns it when you knock it out

At MIT, a robot assembles ten optical components, corrects placement errors and restores laser alignment after a disturbance.

Ten optical components are left in loose positions on an MIT optical table. A robot arm identifies them, picks them up, brings them onto a beam path, builds a laser cavity, turns the tiny screws that steer the mirrors and keeps adjusting until the laser works. Then the researchers deliberately displace a lens on the centimetre scale. The system works out what moved and rebuilds the alignment.[1](https://news.mit.edu/2026/robotic-lab-runs-optics-experiments-on-demand-0917)[2](https://arxiv.org/abs/2603.21496)

The striking part is not that a robot can carry a mirror. Pick-and-place is old industrial territory. What matters is what happens **after** placement: the robot measures the consequences of its own action, notices that the useful optical position was missed, deliberately perturbs the setup to learn how the beam responds, corrects the placement, then switches to a finer actuator when the arm itself is no longer the right tool.[2](https://arxiv.org/abs/2603.21496)

The researchers have done more than automate a sequence of human motions: they have made the experiment legible enough for a machine to judge whether a step actually worked. That requirement matters especially in tabletop free-space optics, where a vial might tolerate an ordinary A-to-B transfer while a mirror sitting close to its nominal coordinate can still ruin the experiment if the beam lands a few hundred micrometres away or leaves at a slightly wrong angle.[2](https://arxiv.org/abs/2603.21496)

A recipe is not enough here. It has to be expressed as coordinates, observables and correction loops.

## Ten parts

The fullest demonstration starts with ten optical elements placed randomly on the table by a person. A user supplies the experimental layout, such as the intended spacing among components. The robot converts that abstract geometry into laboratory coordinates and progressively constructs an optically pumped tabletop laser cavity.[2](https://arxiv.org/abs/2603.21496)

The boundary matters because **the robot is not deciding which scientific experiment should be run**. A layout arrives from the user. The machine's autonomy begins with physically executing, aligning and maintaining it.

The sequence is long enough to expose the difficulty. The system places two feedback cameras and a neutral-density filter, measures the pump beam path, positions the output coupler, and temporarily inserts both a beam block and a beam splitter to establish an alignment reference. It later removes elements, installs a lens, places the input coupler, adds a band-pass filter and finally positions the Nd:YAG crystal that produces the desired laser signal.[2](https://arxiv.org/abs/2603.21496)

At every stage, putting the object down is not the end of the instruction. The optical result decides what happens next.

> CONSTRUCTION
> **A layout becomes physical**
> - The user supplies the intended geometry of the experiment.: 1
> - The robot identifies components and maps them into table coordinates.: 2
> - Each part is placed, then optical feedback exposes the remaining error.: 3
> - Fine adjustment and beam measurements continue until the desired laser mode appears.: 4
> Demonstration: optically pumped laser cavity, Choi et al., IROS 2026.

What first looks like optical Lego is closer to a program in which every instruction checks the state of the world before the next one can safely run.

## The wrong millimetre

The hardware is a seven-degree-of-freedom UFACTORY xArm7 with a 61 cm reach, a 3.5 kg payload and a listed positioning precision of ±0.1 mm.[2](https://arxiv.org/abs/2603.21496) In many industrial tasks, one tenth of a millimetre is already respectable.

Optics can still reject it.

The paper notes that translations of tens to hundreds of micrometres or rotations below a degree can extinguish a useful signal. The same error also matters differently from one component to another: a slightly misplaced filter may continue to work, while a small lens or cavity-mirror offset can stop the system from operating altogether.[2](https://arxiv.org/abs/2603.21496)

That leaves robotics with an awkward stack of errors. The arm is only one contributor: computer vision estimates pose imperfectly, mechanical tolerances accumulate, and the optic was mounted inside its housing by a person, so its optical centre may not perfectly coincide with the housing centre. Rather than demand a perfect first move, the researchers ask the system to **notice when the move was wrong**.

Spatial optimisation begins with the arm placing a component where the geometric model says it belongs. A camera observes where the beam actually lands. The robot moves the component by a known small displacement, measures the corresponding beam shift and uses that response to calculate a correction. It iterates until the beam is close enough to the target.[2](https://arxiv.org/abs/2603.21496)

Precision has moved from the mechanics alone into the **loop** between action and measurement.

The reported numbers make that distinction visible. For the output coupler, the final component-placement precision is ±0.28 mm while beam-centre precision is ±0.14 mm. For a lens, the paper reports ±0.074 mm placement precision and ±0.31 mm on the beam centre.[2](https://arxiv.org/abs/2603.21496) These figures do not describe a magical robot repeating a perfect coordinate. They describe a system correcting until an optical condition is met.

## Three views

Closing that loop does not rely on one all-purpose “smart camera.” The lab distributes perception across three scales.[2](https://arxiv.org/abs/2603.21496)

Stereo 4K cameras watch the table and provide coarse object poses. A LiDAR camera near the end effector helps with accurate grasping. Beam-detection cameras serve as the eyes of optical alignment itself.

No one sensor answers every question.

The global view knows roughly where a housing sits but cannot tell whether two beams overlap. The wrist-level camera improves grasping without measuring laser mode quality, while beam cameras see the optical consequence but would be a terrible way to search an entire table for a misplaced object. The laboratory works, in other words, as a relay among specialised sensing systems rather than as one omniscient robot.

> PERCEPTION
> **Three scales**
> - Stereo 4K cameras estimate poses and monitor the workspace.: Table view
> - An end-effector LiDAR camera supports accurate grasping and placement.: Hand view
> - Dedicated optical cameras measure the consequence of each adjustment.: Beam view
> - Software compares measurement with target and chooses the next correction.: Loop
> Useful precision emerges from the sensor chain rather than a single camera.

This arrangement also deals with a classic optical problem: feedback can be **sparse**. When a mirror is badly aligned, the camera does not necessarily see an obviously bad beam. It may see no beam at all. Useful feedback only exists in a narrow part of the adjustment space.[2](https://arxiv.org/abs/2603.21496)

Automation therefore needs a way to find that region before it can refine anything inside it.

## Give it handles

To make components tractable, the team modifies the laboratory as well as the robot.

Mirrors, lenses and other elements sit inside custom 3D-printed housings with standardised geometry. Each housing carries ArUco fiducial markers for identification, a magnetic base with embedded neodymium magnets and a rubberised underside to reduce slipping.[2](https://arxiv.org/abs/2603.21496)

> Illustration: MIT robotic optics setup and researchers holding a mirror in its standard housing and the motorized fine-adjustment tool. The system does not manipulate arbitrary naked optics. Components receive standard housings, while a dedicated motorized tool operates some of the existing adjustment knobs. Credit: [Jennifer Chu, MIT](https://news.mit.edu/2026/robotic-lab-runs-optics-experiments-on-demand-0917).

This is one of the less glamorous parts of the project, and probably one of the most transferable. An automatable laboratory does not always need a robot dexterous enough to tolerate every object already present. The objects can meet the machine halfway by exposing a stable physical interface. Software would call that an API. Here it has a shape, magnets and a visual marker.

There is a real concession in that choice. A human still performs the one-time job of mounting each optical component in its housing, and that installation can itself introduce a systematic offset. The paper explicitly includes that offset among the errors later absorbed by closed-loop correction.[2](https://arxiv.org/abs/2603.21496)

The system is therefore not universal in the sense of “put any naked optic from any laboratory on the table and it will know what to do.” Its flexibility comes from a bargain: very different components become machine-manipulable because they are given a **common interface**.

That tradeoff echoes a broader problem in self-driving laboratories. Recent field reviews emphasise that transferring autonomous systems across facilities remains hard because hardware is heterogeneous, workflows are bespoke and generalisation requires modular interfaces, interoperable representations and strong experimental provenance.[4](https://www.nature.com/articles/s44182-026-00094-z)[5](https://www.nature.com/articles/s41570-026-00847-2)

MIT has not abolished that problem. It has built an unusually tangible version of the answer.

## The last screw

One operation cannot be solved simply by moving a housing across the table: finely turning the adjustment knobs on kinematic mirror mounts.

The team built a **Fine-Adjustment Tool**, or FAT, a Wi-Fi-controlled motorised device that engages the ordinary knobs already present on standard optical mounts.[2](https://arxiv.org/abs/2603.21496)

The researchers do not motorise every mount permanently. Instead, the robot carries a fine actuator to whichever component currently needs one, leaving the ordinary optical hardware largely intact.

To align an optical resonator, the system has to overlap two beams. It first stores a reference image containing only the primary beam. After the relevant mirror is installed, the camera sees both the primary beam and the secondary beam created by reflections in the cavity. Subtracting the reference isolates the secondary beam, after which Bayesian optimisation turns the mirror knobs to minimise the distance between the two beam centres.[2](https://arxiv.org/abs/2603.21496)

Across ten different initial positions, the procedure overlapped the beams successfully in all ten cases, taking **5.5 ± 0.5 iterations**.[2](https://arxiv.org/abs/2603.21496)

No lookup table says “turn screw A by 0.7 turns.” The system observes the optical response, chooses another adjustment and measures again, which is the point where automating motions becomes automating **criteria**.

## Close the loop

The work follows another project from much of the same group. In August 2026, *Science Advances* published *AI-driven robotics for optics*, a platform that combined generative AI, computer vision and robotics to translate user-defined goals into optical configurations, assemble them with submillimetre accuracy, fine-align them and run measurements such as beam characterisation, polarisation mapping and spectroscopy.[3](https://pubmed.ncbi.nlm.nih.gov/42616899/)

The new IROS work puts less emphasis on generating a layout and much more on what happens when a sensitive physical experiment has to **keep working**.

The authors divide the challenge into three directions: break an optical experiment into smaller automatable tasks, formalise alignment as optimisation, and actively monitor the setup so it can recover after a disturbance.[2](https://arxiv.org/abs/2603.21496)

The third capability changes the category of the machine. Assembly performed once is sophisticated automation. Remembering a useful state, noticing its disappearance and owning a procedure for finding it again begins to look much more like laboratory infrastructure.

Self-driving laboratories in chemistry and materials science already rely heavily on closed loops: choose an experimental condition, make a sample, measure it, analyse the result and use that result to decide what to do next.[4](https://www.nature.com/articles/s44182-026-00094-z) Free-space optics adds an almost irritating prerequisite. Before the scientific question can even be tested, a fragile physical geometry often has to remain aligned merely to keep the instrument capable of measuring.

Here, keeping the beam alive becomes part of autonomy.

## Break it on purpose

After showing that the laser works, the researchers deliberately ruin it.

In the first recovery test, a lens is displaced on the centimetre scale, enough to destroy the expected operation. The end-effector camera compares the current positions of the lens and two critical mirrors with the positions stored before the disturbance. The system identifies the displaced component and attempts to put it back.[2](https://arxiv.org/abs/2603.21496)

One replacement is not always enough. Finite arm precision leaves residual error. If the laser does not return, a realignment routine begins: the robot grips the component, translates it and places it again until the signal is restored.[2](https://arxiv.org/abs/2603.21496)

Across ten trials, the laser signal returned ten times. Only **40%** of trials succeeded with a single pick-and-place. The remaining cases needed an average of **2.7 ± 1.4 attempts** during the extra realignment routine. Average recovery time was 2.83 minutes, close to the 2.6 minutes required for the original lens spatial optimisation.[2](https://arxiv.org/abs/2603.21496)

The headline number “100% success” hides the useful part: the loop exists precisely because the robot is not deterministic enough to assume that returning a component to a stored coordinate will restore the experiment.

> SELF-RECOVERY
> **The first move often misses**
> - Laser signal restored after deliberately displacing the lens.: 10 / 10
> - Recoveries completed with a single replacement.: 40%
> - Realignment attempts in cases that needed an additional loop.: 2.7 ± 1.4
> - Average recovery time measured across the ten trials.: 2.83 min
> Robustness comes from recovery after error, not from perfect placement.

A second test is harsher. All four mirror knobs are randomly rotated by 30° to 60°, completely extinguishing lasing. The system runs Bayesian optimisation until its objective reaches at least 90% of the original value.[2](https://arxiv.org/abs/2603.21496)

This time nine out of ten trials succeed, requiring **12 ± 9 iterations** and 3.05 minutes on average.[2](https://arxiv.org/abs/2603.21496)

The tenth failure matters because it blocks the easy story about total autonomy. The system recovered from several severe disturbances in this setup, but an arbitrary optical experiment has not been shown to repair its own alignment.

## Pick the mode

A laser cavity can also “work” without working in the desired way.

Small alignment errors encourage higher-order modes: a signal exists, but its spatial shape is not the desired fundamental Gaussian mode. The researchers use the beam-quality factor M², where 1 corresponds to a diffraction-limited Gaussian beam, to distinguish these states.[2](https://arxiv.org/abs/2603.21496)

The optimisation protocol begins with a coarse crystal-angle sweep in **0.2° increments**. In the example shown, a laser signal appears only inside a narrow region around 1.2° to 1.4°. Once the robot locates that region, it releases the crystal and hands fine alignment to the motorised mirror tool. The objective combines beam intensity with M², so the system searches for power and mode quality together.[2](https://arxiv.org/abs/2603.21496)

That sequence captures the project in miniature: a coarse search with the whole arm finds a region where the physical phenomenon becomes observable, after which the system switches both actuator and metric for the fine search. The human protocol is no longer a motion-by-motion script. It has become a sequence of measurable objectives.

## Not the scientist

MIT says the group is developing a cloud application that could eventually let researchers access physical robotic labs remotely and submit experimental protocols or queries.[1](https://news.mit.edu/2026/robotic-lab-runs-optics-experiments-on-demand-0917) The IROS paper also suggests predictive simulation, learning-based methods and agentic AI as future additions that could simplify experiment design and execution.[2](https://arxiv.org/abs/2603.21496)

Those are **future directions**, not demonstrated capabilities here.

In the laser-cavity experiment, a person supplies the layout. The optics have been prepared in robot-friendly housings. The system knows the required tasks and uses objectives designed for this setup. Its most quantified recovery tests cover ten starting conditions or ten disturbances.[2](https://arxiv.org/abs/2603.21496)

None of that makes the result small. Building a laser cavity from scattered parts, aligning its beams and restoring the signal without human intervention during the sequence is still a demanding physical demonstration.

But “autonomous laboratory” hides two different kinds of autonomy.

> AUTONOMY
> **Two different questions**
> - Build, measure, correct, tune and restore an experiment whose layout is known.: Execute
> - Choose the next hypothesis, invent the protocol and redefine the scientific objective.: Decide
> - The IROS system automates much of the closed-loop physical execution.: Shown here
> - Agentic AI, predictive simulation and cloud access are described as later extensions.: Future work
> Automating a protocol is not the same thing as autonomously choosing the science.

The first question is: “Can you run this experiment correctly without me?”
The second is: “Can you decide which experiment deserves to be run next?”

This work mainly answers the first, a distinction that also appears in recent writing on self-driving labs. The field is increasingly concerned with scalability, transfer between laboratories and end-to-end provenance rather than treating “AI + robot” as a sufficient definition of autonomy.[4](https://www.nature.com/articles/s44182-026-00094-z)[5](https://www.nature.com/articles/s41570-026-00847-2)

MIT contributes a particularly important missing layer for optics: how to structure a reconfigurable setup enough for a machine to build it while instrumenting the process enough for the machine to know that the build still is not good enough.

## The lab as software

The most useful detail may be the one least visible in a demo: when a person runs an optical experiment, a large part of the protocol lives in their body. “Put the mirror here” also means: watch where the beam lands, approach gently, turn this knob, reverse if the signal disappears, search the region where it comes back and refine until the right pattern appears.

A lab notebook can describe those steps in prose. It rarely captures the full recovery strategy, practical tolerances and changes of metric that make an experienced operator effective.

The robot forces those implicit rules into executable form. A component needs a machine-readable identity, its position has to be observable and each stage needs its own success criterion. An error must trigger something more precise than a human frown, and history has to be stored if a previous working state is going to be reconstructed.

That is why the project is more interesting than a photogenic arm over an optical table: the machine automates the hand only after the team has **formalised the craft surrounding the hand**.

Software has lived with a version of this idea for decades: a process becomes truly programmable when its inputs, states, errors and exit conditions are explicit enough to replay.

In this laboratory, ArUco markers look a little like addresses. Standard housings become interfaces. Cameras expose current state. The beam behaves like a test that can fail. Spatial and angular optimisation are correction loops. The Fine-Adjustment Tool looks like an adapter for a legacy human interface, namely the tiny screw nobody originally designed for a robot.

The analogy has a useful limit because physical reality keeps slipping, vibrating, heating and drifting. A mirror cannot be “rerun” exactly like a function. That friction is the point. The protocol becomes programmable by assuming from the start that matter **will not do exactly what it was told on the first attempt**.

And in nine of ten tests where the researchers scramble the four adjustments keeping the laser aligned, the lab finds its own route back to a working beam.[2](https://arxiv.org/abs/2603.21496)

## References

1. [MIT News, “Robotic lab sets up and runs optics experiments on demand,” September 17, 2026](https://news.mit.edu/2026/robotic-lab-runs-optics-experiments-on-demand-0917)
2. [Seou Choi et al., “A Framework for Closed-Loop Robotic Assembly, Alignment and Self-Recovery of Precision Optical Systems,” arXiv:2603.21496, IROS 2026](https://arxiv.org/abs/2603.21496)
3. [Shiekh Zia Uddin et al., “AI-driven robotics for optics,” Science Advances 12(34), 2026](https://pubmed.ncbi.nlm.nih.gov/42616899/)
4. [Christopher H. J. Moran, Jinge Xu & Milad Abolhasani, “The rise of robotics in self-driving labs,” npj Robotics 4, 29, 2026](https://www.nature.com/articles/s44182-026-00094-z)
5. [Richard B. Canty & Milad Abolhasani, “The past, present and future of self-driving laboratories,” Nature Reviews Chemistry 10, 523–537, 2026](https://www.nature.com/articles/s41570-026-00847-2)